développement d`une méthode de boltzmann sur réseau pour la

Transcription

développement d`une méthode de boltzmann sur réseau pour la
UNIVERSITÉ DE MONTRÉAL
DÉVELOPPEMENT D’UNE MÉTHODE DE BOLTZMANN SUR RÉSEAU
POUR LA SIMULATION D’ÉCOULEMENTS VISQUEUX EN CUVES
AGITÉES
VINCENT STOBIAC
DÉPARTEMENT DE GÉNIE CHIMIQUE
ÉCOLE POLYTECHNIQUE DE MONTRÉAL
THÈSE PRÉSENTÉE EN VUE DE L’OBTENTION
DU DIPLÔME DE PHILOSOPHIÆ DOCTOR
(GÉNIE CHIMIQUE)
JANVIER 2013
© Vincent Stobiac, 2013.
UNIVERSITÉ DE MONTRÉAL
ÉCOLE POLYTECHNIQUE DE MONTRÉAL
Cette thèse intitulée:
DÉVELOPPEMENT D’UNE MÉTHODE DE BOLTZMANN SUR RÉSEAU POUR LA
SIMULATION D’ÉCOULEMENTS VISQUEUX EN CUVES AGITÉES
présentée par: STOBIAC Vincent
en vue de l’obtention du diplôme de : Philosophiæ Doctor
a été dûment acceptée par le jury d’examen constitué de :
M. PERRIER Michel, Ph. D., président
M. BERTRAND François, Ph. D., membre et directeur de recherche
M. TANGUY Philippe A., Ph. D., membre et codirecteur de recherche
M. DUFOUR Steven, Ph. D., membre
M. DERKSEN Jos, Ph. D., membre
iii
CITATION
Tout petit, je voulais être célèbre et je ne faisais rien pour. À l’école, je m’avérais très vite un
élève inexistant. Par goût. J’ai toujours été persuadé – je le suis encore – que les diplômes sont
faits pour les gens qui n’ont pas de talent. Malheureusement, il ne suffit pas de ne pas avoir de
diplômes pour avoir du talent.
Pierre Desproges – Chroniques de la haine ordinaire
iv
REMERCIEMENTS
Je tiens tout d’abord à remercier mon directeur et mon codirecteur de thèse, les professeurs
François Bertrand et Philippe Tanguy, pour m’avoir accordé leur confiance et donné
l’opportunité de faire ce projet. Sans leur vision éclairée, leurs précieux conseils et leurs
encouragements, je n’aurais jamais pu réaliser ce travail.
Je tiens aussi à remercier les membres du jury, les professeurs Michel Perrier, Jos Derksen,
Steven Dufour et Marcelo Reggio, pour avoir accepté d’évaluer mon travail.
Parmi les nombreuses personnes qui ont contribué à ma formation scientifique, je tiens à
remercier très chaleureusement Mourad Heniche. Sans sa patience et sa gentillesse, je n’aurais
jamais pu acquérir l’autonomie nécessaire à la réalisation d’un tel projet. Je tiens aussi à
remercier le professeur Louis Fradette pour ses critiques constructives et son oreille attentive à
mes élucubrations numériciennes.
Je remercie aussi François Drolet et David Vidal, de FPInnovations, pour les nombreuses
discussions enrichissantes que nous avons eues. Il est important de souligner que les résultats de
l’Article 1 ont été obtenus grâce au code développé par David.
Je tiens aussi à remercier Jean-Sébastien Cagnone, Christophe Devals et Cyrille Séguineau de
Préval de m’avoir aidé à pousser plus loin mes réflexions.
Bien entendu, cette expérience aurait été beaucoup moins sympathique sans la présence d’André,
Bruno, Florent, Florian, Jonathan, Julie, Mehdi, Olivier, Paul et Yoann à l’URPEI.
Je tiens aussi à remercier mes parents de m’avoir soutenu durant toute ma scolarité. Sans leurs
encouragements et leur support, je n’en serais pas là aujourd’hui.
Finalement, je remercie l’École Polytechnique de Montréal et le département de Génie chimique
de m’avoir accueilli. Je tiens aussi à remercier le Canada et le Québec pour cette expérience
personnelle très enrichissante.
v
RÉSUMÉ
La caractéristique principale des écoulements visqueux est leur forte résistance à l’écoulement.
Comme la prédominance des forces visqueuses empêche la génération de turbulence, la forme de
l’écoulement est principalement affectée par la géométrie du problème, la rhéologie du fluide et
la présence de plusieurs phases. La simulation de tels écoulements requiert donc l’utilisation d’un
outil numérique capable de traiter précisément les géométries, les rhéologies et les interactions
entre les phases. Dans la cadre de cette thèse, l’étude des méthodes numériques a pour objectif
d’améliorer la simulation des écoulements visqueux en cuves agitées.
La simulation des écoulements dans les mélangeurs permet de compléter les informations
obtenues lors des mesures expérimentales. Par exemple, les approches numériques permettent de
visualiser facilement les écoulements à l’intérieur du système de mélange, ainsi que de
caractériser des grandeurs globales comme le pompage. Pour autant, les résultats ne sont
intéressants que si l’on maitrise l’impact des différents paramètres numériques sur la précision.
Le cas des mélangeurs industriels spécialisés pour les liquides visqueux concentre un très grand
nombre de difficultés. Premièrement, la géométrie du système de mélange est complexe et évolue
au cours du temps. En effet, ce type de mélangeur utilise des formes de pales complexes afin de
réaliser le mélange. De plus, les pales doivent racler la cuve pour prévenir la stagnation du
liquide, ce qui crée des espacements faibles comparativement à la taille du système.
Deuxièmement, les propriétés physiques du mélange sont généralement compliquées. La
difficulté réside ici dans la modélisation de l’interaction entre les phases ou entre les différents
liquides. Dans le cas de l’étude de l’agitation, caractérisée par la simulation d’une seule phase
liquide homogène, la difficulté se concentre sur la rhéologie du liquide. Dans le cadre de cette
thèse, nous nous limiterons à l’étude de l’agitation.
Malgré ces difficultés, plusieurs travaux de recherche ont été effectués sur la simulation du
mélange visqueux dans des cuves agitées utilisées dans l’industrie. Il ressort de ces articles que ce
sont principalement les méthodes d’éléments et de volumes finis qui sont utilisées. Malgré des
simplifications astucieuses pour faciliter le traitement de la géométrie (tels que les changements
de repère ou les conditions aux limites immergées), la simulation des systèmes de mélange
entraine un compromis entre la précision des résultats et les temps de calcul. Cette contrainte peut
devenir un facteur limitant pour des géométries très complexes et/ou des rhéologies largement
vi
non newtoniennes. Cela s’explique principalement par la taille et la complexité du système
d’équations obtenu. Ainsi, dans la littérature, les travaux se situant à l’échelle industrielle se
limitent généralement à une seule phase liquide. Certaines simulations polyphasiques sont
pourtant réalisées, mais souffrent de temps de calcul très importants et/ou d’approximations
affectant globalement la précision. Afin de dépasser ces contraintes, il semble pertinent d’utiliser
une méthode efficace pour le calcul parallèle et qui permet de traiter facilement des géométries et
des rhéologies complexes. C’est pour cette raison que la méthode de Boltzmann sur réseau est
utilisée dans cette thèse.
La méthode de Boltzmann sur réseau « lattice Boltzmann method (LBM) », basée sur la théorie
cinétique des gaz, part du principe que la simulation du mouvement de particules de fluide à une
échelle physique plus petite (mésoscopique), va permettre d’obtenir le comportement
macroscopique du fluide. Ainsi, à partir de l’équation de Boltzmann, la solution des équations de
Navier-Stokes peut être obtenue. L’avantage de ce schéma numérique est qu’il est explicite et
basé sur des calculs simples et locaux, ce qui lui confère une très bonne efficacité pour le calcul
parallèle. À l’inverse, ce changement d’échelle physique comporte aussi des inconvénients,
comme la difficulté d’imposer des conditions aux limites macroscopiques. Malgré des
approximations sur les phénomènes aux frontières, la LBM a été appliquée avec succès pour la
simulation d’écoulements en milieux poreux et de suspensions solide-liquide.
L’objectif principal de cette thèse est de développer une méthode de Boltzmann sur réseau
efficace pour la simulation d’écoulements visqueux en cuves agitées. Pour ce faire, il est
nécessaire d’étudier l’applicabilité des différentes stratégies pour l’imposition des conditions aux
limites et de caractériser la précision de la LBM dans le cas de géométries complexes.
Finalement, il est aussi intéressant d’étudier l’applicabilité de la LBM pour des rhéologies
complexes.
L’article 1 montre que, dans le cas de géométries courbes (Couette cylindrique et mélangeurs), la
précision du schéma LBM est largement affectée par le choix de la stratégie d’imposition de
conditions aux limites. Par exemple, dans le cas des mélangeurs raclants, l’approche dite des
conditions aux limites immergées « immersed boundary method (IBM) » apparait comme étant
moins adaptée. À l’inverse, l’approche la plus simple « modified bounce back-rule (MBB) »,
basée sur la règle du rebond, et l’approche plus élaborée « extrapolation method (EM) », qui
vii
extrapole les populations en fonction de la distance à l’obstacle, donnent la même précision sur
les grandeurs caractéristiques des mélangeurs. Cependant, comme seule la stratégie EM permet
d’étudier précisément l’impact d’une petite variation de la géométrie, cette approche est
sélectionnée pour la suite des travaux.
L’article 2 montre ensuite que la LBM, couplée avec la méthode d’extrapolation, permet
d’obtenir une précision adéquate pour la simulation de la consommation de puissance et de la
capacité de pompage du système de mélange Maxblend. Elle permet aussi d’étudier l’impact de
paramètres géométriques comme l’espacement entre la pale et le fond de la cuve. De plus, dans le
cas d’une géométrie variable dans le temps (configuration avec contre-pales), la précision reste
bonne même si les vitesses sont perturbées aux alentours de la pale.
Finalement, l’article 3 montre que, sous certaines limitations sur nombre de Reynolds, la LBM
peut être utilisée efficacement pour la simulation d’écoulements fortement non newtoniens.
Ainsi, dans le cas du système de mélange Maxblend, un modèle rhéologique de type loi de
Carreau-Yasuda avec un indice de puissance de 0.05 a pu être simulé sans affecter largement les
temps de calcul.
En conclusion, cette thèse a permis de montrer que la LBM peut être utilisée efficacement pour la
simulation des écoulements dans les mélangeurs industriels spécialisés pour les liquides
visqueux. Il a même été montré que cette approche pouvait être utilisée pour faire de
l’optimisation géométrique des systèmes de mélange. De plus, sous certaines contraintes relatives
au régime d’écoulement, il a aussi été montré que la LBM permettait de simuler l’agitation de
liquides avec des comportements fortement non newtoniens.
viii
ABSTRACT
The major feature of the viscous flow is its resistance to motion. As the viscous forces are
dominant the turbulent phenomena disappear, whereby the shape of the flow is determined by the
geometry of the problem, the rheology of the fluid and the presence of several phases. In order to
simulate such flows, the selected numerical tool used should be capable of accurately managing
the geometry, the rheology and interactions between phases. In the scope of this thesis, numerical
methods are assessed in order to improve the simulation of viscous flows in stirred tanks.
In the case of mixing systems, computational fluid dynamics (CFD) is a facilitating tool which
may be used to supplement experimental measurements. For instance, numerical simulations
allow to readily visualize the flow inside the mixer and determine characteristic mixing numbers,
such as the pumping capacity. However, simulation results are useful only if the impact of the
numerical parameters on the accuracy is understood. Batch mixers for highly viscous fluids are
quite tricky, rendering the simulation process highly challenging. First, the geometry of the
mixing system is complex and time dependent. The complex shape of the geometry is though
necessary and used to compensate the impossibility of achieving adequate mixing with the
turbulence. Given that close clearance impellers are used to prevent the fluid stagnation near the
tank, small gaps are found between the tank and impeller. Second, the physical properties of the
mixture are also complex. The modelling approach should be able to account for the interaction
between the phases (solid, liquid or gas). In the case of blending, only one single liquid phase is
considered, thus the difficulty is concentrated on the rheology of the fluid. In this thesis, only
blending will be examined.
Notwithstanding such difficulties, the simulation of viscous mixing in stirred tanks has been the
subject of many articles. It can be inferred from the literature that the finite element method
(FEM) and finite volume method (FVM) have been the most commonly used approaches.
Despite the use of ingenious ways to treat the complexity of the geometry (change of reference
frame or immersed boundary conditions), the simulation of industrial mixing flows leads to a
compromise between accuracy and computation time. This can be explained by the size and
complexity of the system of equations to be solved. This limitation can be rather restrictive in the
case of particularly complex geometries or highly non-Newtonian behaviors. For instance, most
simulations of industrial mixing process consider only one single liquid phase. Conversely, few
ix
multiphase simulations have been applied and suffer from considerable computational cost and/or
approximations affecting the accuracy of the simulation. In order to go beyond these limitations,
it seems pertinent to use a numerical approach efficient for massive parallel computations. In
addition, this method must be able to easily integrate complex geometries and rheology. For the
above reasons, the lattice Boltzmann method has been selected in this work.
The lattice Boltzmann method (LBM) is based on the kinetic theory of gas. The idea behind this
approach is to obtain the macroscopic dynamics of a fluid from the simulated motion of
mesoscopic fluid particles. The solution of the Navier-Stokes equations is then recovered from
the Boltzmann equation. As the LBM is explicit and only requires simple and local computations,
the resulting scheme is highly efficient for parallel computations. Inversely, the change in
physical scale behind the LBM leads to several difficulties, such as the imposition of
macroscopic boundary conditions. Despite approximations on the phenomena at the boundaries,
the LBM has been successfully applied for the simulation of porous media flows and solid-fluid
suspensions.
The main objective of this thesis is to develop an efficient lattice Boltzmann method for the
simulation of viscous flows in stirred tanks. To do so, the applicability of the different strategies
to impose boundary conditions is first compared. Second, the accuracy of the LBM in the case of
complex geometry is characterized. Lastly, the applicability of the LBM for highly nonNewtonian fluid is investigated.
Article 1 illustrates that the accuracy of the scheme is largely affected by the selection of the
boundary condition strategy. For instance, in the case of close clearance impellers, the immersed
boundary method (IBM) seems to be less accurate. Inversely, the simplest approach, the modified
bounce-back rule (MBB), and a more elaborate method, the extrapolation method (EM), which
uses the distance from the obstacle to extrapolate missing populations, provide a similar accuracy
on characteristic mixing numbers. However, only the EM allows investigating variations of the
geometry lower than the lattice spacing. For this reason, this boundary condition strategy is used
in the other articles.
Article 2 shows that the LBM accurately predicts the power consumption and pumping capacity
of a mixing system. In addition, the impact of geometrical parameters, such as bottom clearance,
x
can likewise be effectively investigated. Moreover, in the case of time dependent geometries, the
LBM conserves an acceptable accuracy even if perturbations appear near the impeller.
Lastly, article 3 demonstrates that, under constraint on the maximum Reynolds number, the LBM
can be used to simulate highly non-Newtonian fluid flows. For instance, a Carreau-Yasuda model
with a power index of 0.05 was used without largely affecting the computational cost of the
simulation.
In conclusion, this thesis has permitted to verify the degree of efficiency and utilization of the
LBM for the simulation of viscous flows in stirred tanks. It has also shown that this approach can
be used for the optimisation of process geometry. Moreover, under some limitations of the
Reynolds number, it can be extended to the blending of highly non-Newtonian fluids.
xi
TABLE DES MATIÈRES
CITATION .................................................................................................................................... III
REMERCIEMENTS ..................................................................................................................... IV
RÉSUMÉ........................................................................................................................................ V
ABSTRACT ............................................................................................................................... VIII
TABLE DES MATIÈRES ............................................................................................................ XI
LISTE DES TABLEAUX ........................................................................................................... XV
LISTE DES FIGURES ............................................................................................................... XVI
LISTE DES SIGLES ET ABRÉVIATIONS ............................................................................. XXI
LISTE DES ANNEXES ....................................................................................................... XXVIII
CHAPITRE 1
INTRODUCTION ............................................................................................... 1
1.1
Contexte ........................................................................................................................... 1
1.2
Les écoulements visqueux complexes.............................................................................. 2
1.2.1 Les écoulements dans des géométries complexes ........................................................ 2
1.2.2 Les modèles rhéologiques ............................................................................................ 3
1.2.3 Les écoulements polyphasiques ................................................................................... 5
1.3
Le mélange visqueux en cuves agitées ............................................................................. 6
1.3.1 Les géométries des mélangeurs .................................................................................... 7
1.3.2 Les grandeurs caractéristiques des mélangeurs ............................................................ 8
1.3.3 La simulation du mélange visqueux ........................................................................... 11
1.4
Objectif principal de la thèse .......................................................................................... 15
CHAPITRE 2
REVUE DE LITTÉRATURE : MÉTHODE DE BOLTZMANN SUR RÉSEAU
........................................................................................................................... 16
2.1
De Boltzmann à Navier-Stokes ...................................................................................... 16
xii
2.1.1 Équation de Boltzmann sous forme continue ............................................................. 16
2.1.2 Équation de fermeture ................................................................................................ 18
2.1.3 Discrétisation en vitesse ............................................................................................. 20
2.1.4 Discrétisation spatiale et temporelle .......................................................................... 22
2.2
Équation de collision-propagation ................................................................................. 24
2.3
Définition des conditions aux limites ............................................................................. 26
2.3.1
La règle du rebond...................................................................................................... 27
2.3.2 Stratégies d’interpolation ou d’extrapolation ............................................................. 28
2.3.3 Conditions aux limites immergées ............................................................................. 29
2.4
Mouvement d’une partie mobile .................................................................................... 31
2.5
Intégration des modèles rhéologiques ............................................................................ 32
2.6
La LBM appliquée aux cuves agitées ............................................................................ 34
CHAPITRE 3
CHAPITRE 4
ORGANISATION DE LA THÈSE................................................................... 35
COMPARAISON
DES
DIFFÉRENTES
STRATÉGIES
POUR
L’IMPOSITION DES CONDITIONS AUX LIMITES EN LBM................................................ 38
4.1
Présentation du premier article ....................................................................................... 38
4.2
Boundary conditions for the lattice Boltzmann method in the case of viscous mixing
flows ........................................................................................................................................ 39
4.2.1 Summary .................................................................................................................... 39
4.2.2 Introduction ................................................................................................................ 39
4.2.3 The lattice Boltzmann method ................................................................................... 44
4.2.4 Boundary condition strategies .................................................................................... 48
4.2.5 Convergence study ..................................................................................................... 61
4.2.6 Mixing simulations ..................................................................................................... 71
4.2.7 Parallel performance and memory usage ................................................................... 80
xiii
4.2.8 Conclusion .................................................................................................................. 84
4.2.9 Acknowledgments ...................................................................................................... 86
CHAPITRE 5
ÉTUDE DE LA PRÉCISION DE LA LBM DANS LE CAS DE
GÉOMÉTRIES COMPLEXES ..................................................................................................... 87
5.1
Présentation du deuxième article.................................................................................... 87
5.2
A thorough investigation of the accuracy of the extrapolation method for the lattice
Boltzmann simulation of viscous fluid flow in a Maxblend impeller system ............................ 88
5.2.1 Summary .................................................................................................................... 88
5.2.2 Introduction ................................................................................................................ 88
5.2.3 The lattice Boltzmann method ................................................................................... 91
5.2.4 Analysis of the spatial convergence ........................................................................... 96
5.2.5 Simulation results for the Maxblend mixer .............................................................. 100
5.2.6 Concluding remarks ................................................................................................. 112
5.2.7 Acknowledgments .................................................................................................... 113
5.3
Résultats complémentaires ........................................................................................... 113
CHAPITRE 6
APPLICABILITÉ DE LA LBM POUR LES RHÉOLOGIES COMPLEXES.
..................................................................................................................... 117
6.1
Présentation du troisième article .................................................................................. 117
6.2
Pumping characterization of the Maxblend impeller for Newtonian and strongly non-
Newtonian fluids ...................................................................................................................... 118
6.2.1 Summary .................................................................................................................. 118
6.2.2 Introduction .............................................................................................................. 118
6.2.3 Methodology ............................................................................................................ 120
6.2.4 Simulation results ..................................................................................................... 129
6.2.5 Conclusions .............................................................................................................. 143
xiv
6.2.6 Acknowledgments .................................................................................................... 144
6.3
Résultats complémentaires ........................................................................................... 144
CHAPITRE 7
DISCUSSION GÉNÉRALE ........................................................................... 147
CONCLUSION ET RECOMMANDATIONS ........................................................................... 155
BIBLIOGRAPHIE ...................................................................................................................... 158
ANNEXES .................................................................................................................................. 170
xv
LISTE DES TABLEAUX
Tableau 2.1: Directions
et coefficients
pour la lattice D3Q15. .......................................... 22
Tableau 4.1: Total number of fluid (F) and solid (S) lattice nodes for the Paravisc impeller
system. .................................................................................................................................... 76
Tableau 4.2: Characteristics of the supercomputer used for the simulation. ................................. 82
Tableau 4.3: Memory usage for the three boundary conditions strategies. .................................... 83
Tableau 5.1: Formulas for the parameters of the extrapolation method (EM). .............................. 94
Tableau 5.2: Geometrical characteristics of the Maxblend mixing system and physical properties
of the fluid. ........................................................................................................................... 101
Tableau 5.3: Characteristics of the supercomputers used for the simulations. ............................ 103
Tableau 5.4: Values of the power constant for the standard bottom clearance obtained with the
LBM, with the FEM [13] and measured experimentally [143]............................................ 104
Tableau 6.1: Geometrical characteristics of the Maxblend mixing system. ................................ 121
Tableau 6.2: Values of the rheological parameters for the Newtonian and non-Newtonian fluids.
.............................................................................................................................................. 122
Tableau 6.3: Characteristics of the Colosse supercomputer. ....................................................... 128
xvi
LISTE DES FIGURES
Figure 1.1: Espace poreux d'un morceau de grès [4]. ...................................................................... 2
Figure 1.2: Simulation LBM dans un pot catalytique (tirée de [7]) ................................................. 3
Figure 1.3: Illustrations des comportements rhéologiques............................................................... 5
Figure 1.4: Discrétisation d'une suspension (simulation 2D) [19] ................................................... 5
Figure 1.5 : Exemples de mélangeurs spécialisés pour les liquides visqueux : (a) mélangeur
statique Kenics, (b) mélangeur planétaire tripale et (c) double ruban hélicoïdal. .................... 7
Figure 1.6: Mélangeur Maxblend (SHI, 2008) ................................................................................. 8
Figure 1.7: Courbes de temps de mélange pour différents systèmes d'agitation (données
provenant de Yamato et al., 1998 [34]) .................................................................................... 9
Figure 1.8: Courbes de puissance pour le mélangeur Maxblend (données provenant de
Guntzburger et al., 2012 [35]). ............................................................................................... 10
Figure 1.9: Illustrations des différentes approches de gestion de la géométrie. (a) changement de
repère, (b) utilisation de deux repères et (c) conditions aux limites immergées.
et
réfèrent aux repères fixes par rapport à la cuve et à la pale, respectivement. ........................ 12
Figure 2.1: Représentation de la lattice D3Q15 ............................................................................. 22
Figure 2.2: Illustration de la règle du rebond (tirée de l'article 1) .................................................. 27
Figure 2.3: Illustration de la distance à la frontière (tirée de l'article 1) ........................................ 28
Figure 2.4: Illustration d'une condition aux limites immergée ...................................................... 30
Figure 2.5: Illustration de nouvelles lattices fluides après mouvement d'un objet ........................ 31
Figure 4.1: Populations used for the half-way bounce-back rule ................................................... 49
Figure 4.2: Populations used for MBB........................................................................................... 50
Figure 4.3: Discretization with MBB ............................................................................................. 51
Figure 4.4: Discretization with EM ................................................................................................ 51
Figure 4.5: Fraction to the boundary with EM ............................................................................... 52
xvii
Figure 4.6: Example of control point distribution with IB-LBM ................................................... 56
Figure 4.7: Couette flow geometry ................................................................................................ 61
Figure 4.8: Convergence study for the Couette flow problem with (a) the MBB, (b) the EM and
(c) the IB-LBM....................................................................................................................... 65
Figure 4.9: Angular velocity profiles obtained with (a) the MBB, (b) the EM and the (c) IB-LBM.
The grey line represents the analytical solution ..................................................................... 66
Figure 4.10: Average local error profiles with (a) the MBB, (b) the EM and (c) the IB-LBM ..... 67
Figure 4.11: Error study on the extrapolation parameter of the EM [(F) refers to the use of the
free parameter; (C) refers to the regular definition] ............................................................... 67
Figure 4.12: Variation of the error versus the mesh spacing for the EM (the grey line refers to the
analytical fraction, the circles refer to fraction computed using a mesh) .............................. 68
Figure 4.13: Impact of the discretization of the inner and outer cylinders on the accuracy for the
EM (the squares refer to the analytical fraction, the circles refer to fraction computed using a
mesh) ...................................................................................................................................... 69
Figure 4.14: Impact of the control point distribution on the error for the IB-LBM ....................... 70
Figure 4.15: Impact of the number of control points on the order of convergence of the IB-LBM
................................................................................................................................................ 71
Figure 4.16: Geometry of the anchor impeller (left) and the Paravisc (Ekato GmbH) impeller
(right). ..................................................................................................................................... 71
Figure 4.17: Power number for the anchor impeller system: (a) the MBB, (b) the EM and (c) the
IB-LBM .................................................................................................................................. 77
Figure 4.18: Convergence of Kp with respect to the lattice spacing for the anchor impeller system
................................................................................................................................................ 78
Figure 4.19: Convergence of Kp with respect to the lattice spacing for the Paravisc impeller
system ..................................................................................................................................... 79
Figure 4.20: Convergence of the axial flow number
with respect to the lattice spacing for the
Paravisc impeller system ........................................................................................................ 79
xviii
Figure 4.21: Contours of the norm of the velocity on a horizontal cross-section with (a) the MBB,
(b) the EM and the (c) IB-LBM. (
,
and
................ 80
Figure 4.22: Graph of the computational time vs the number of processors ................................. 82
Figure 5.1: Distance to the boundary with the extrapolation method (EM). ................................. 93
Figure 5.2: Geometry of the Couette flow problem. ...................................................................... 96
Figure 5.3: Convergence curves for the Couette problem with periodic boundary conditions: (a)
velocity and shear rate, (b) power consumption from shear rate and torque values. The
numbers correspond to the slopes of the straight lines........................................................... 98
Figure 5.4: Convergence curves for the Couette flow problem with mixed boundary conditions:
(a) velocity and shear rate, (b) power consumption and (c) sum over the lattice nodes of the
axial component of the velocity,
. The numbers correspond to the slopes of the straight
lines. ....................................................................................................................................... 99
Figure 5.5 : The Maxblend impeller (wedge configuration): (a) geometry and (b) illustration of
the velocity profile at Re=120. ............................................................................................. 100
Figure 5.6: Comparison of LBM, FEM [13] and experimental data [143] for (a) the impact of the
bottom clearance on the power constant and (b) the impact of the Reynolds number on the
axial flow number................................................................................................................. 105
Figure 5.7: Impact of the bottom clearance on (a) the power number, (b) the axial flow number
and (c) their ratio, for Reynolds numbers of 7, 33 and 68. .................................................. 107
Figure 5.8: Evolution of the power number for the unbaffled configuration at Re=7 (
. .............................................................................................................................. 109
Figure 5.9: Variations with respect to the lattice spacing of (a) the average power number and (b)
the standard deviation of the power number, for the unbaffled configuration at Re=7. ...... 110
Figure 5.10: Evolution of the power number at Re=120 for (a) the unbaffled and (b) baffled
configurations, with
. ..................................................................................... 111
Figure 5.11: Variation of the power for the baffled configurations: (a) increase compared to the
unbaffled configuration and (b) maximum amplitude during one revolution. ..................... 111
xix
Figure 5.12 : Impact du facteur de relaxation
sur la précision de (a) la vitesse et (b) du taux de
cisaillement pour le problème de Couette. ........................................................................... 114
Figure 5.13: Impact du facteur de relaxation
sur la précision de la puissance calculée à partir (a)
du taux de cisaillement et (b) du couple pour le problème de Couette. ............................... 115
Figure 5.14: Erreur sur le volume en fonction de la taille de lattice pour le problème de Couette
cylindrique. ........................................................................................................................... 116
Figure 6.1: Geometry of the Maxblend mixing system. .............................................................. 121
Figure 6.2: Shear viscosity data for aqueous solutions of Steol with different salt concentrations
from 0.4 to 2.9 wt% (from [154])......................................................................................... 123
Figure 6.3: Power and pumping curves for the Maxblend mixing system in the Newtonian case
with experimental data [35], FEM [13] and LBM results. ................................................... 130
Figure 6.4: Magnitudes (Vrz) of the radial and axial components of the velocity for the
Newtonian fluid obtained with the LBM, PIV [154] and FEM [13].................................... 132
Figure 6.5: Dimensionless axial velocity profiles at z/H=0.75 with the LBM, PIV [154] and FEM
[13]. ...................................................................................................................................... 133
Figure 6.6: Axial pumping volumes of the Newtonian flow predicted with the LBM, at Reynolds
numbers of 2 and 23: (a) and (c) correspond to three-dimensional views and (b) and (d) to
top views. The red and blue colors denote positive and negative values of the axial
component of the velocity, respectively. The dashed lines represent the section used for the
PIV measurements, and the arrows indicate the sense of rotation. ...................................... 135
Figure 6.7: Variation of the relative importance of three components of the time-average mean
velocity obtained with the LBM in the Newtonian case. ..................................................... 136
Figure 6.8: Power and pumping curves for the Maxblend mixing system: experimental data [35]
and LBM results for the (N) Newtonian and (NN) non-Newtonian cases. .......................... 138
Figure 6.9: Segregation zones for
. Adapted from [35]. ................................................ 138
Figure 6.10: Magnitudes (Vrz) of the radial and axial components of the velocity for the strongly
non-Newtonian fluid obtained with the LBM and PIV [154] for (a)
(b)
(fluid 3) and
(fluid 4)............................................................................................................ 139
xx
Figure 6.11: Axial pumping volumes of the strongly non-Newtonian flow predicted with the
LBM at
of (a) 1, (b) 15, (c) 30 and (d) 50. The red and the blue colors denote positive
and negative values of the axial component of the velocity, respectively. .......................... 140
Figure 6.12: Viscosity maps obtained with the LBM for the strongly non-Newtonian fluid at
of 1, 15, 30 and 50. ............................................................................................................... 142
Figure 6.13: (a) Variation of the relative importance of the three components of the time-average
mean flow for the non-Newtonian case and (b) comparison of the axial components for the
Newtonian and non-Newtonian fluids.................................................................................. 143
Figure 6.14: Impact de l’indice de puissance
sur la précision de (a) la vitesse et (b) du taux de
cisaillement pour le problème de Couette. ........................................................................... 145
Figure 6.15: Impact de l’indice de puissance
sur la puissance calculée à partir (a) du taux de
cisaillement et (b) du couple pour le problème de Couette. ................................................. 146
xxi
LISTE DES SIGLES ET ABRÉVIATIONS
Abréviations
BC
Conditions aux limites « Boundary conditions »
BGK
Modèle de collision de Bhatnagar, Gross et Krook
CFD
Calcul de la dynamique des fluides « Computational fluid dynamics »
DHR
Ruban hélicoïdal double « Double helical ribbon »
DNS
Simulation numérique directe « Direct numerical simulation »
DPM
Modèle des particules discrètes « Discrete particle method »
EM
Méthode d’extrapolation « Extrapolation method »
FEM
Méthode des éléments finis « Finite element method »
FVM
Méthode des volumes finis « Finite volume method »
IB-LBM
Conditions immergées pour la méthode de Boltzmann sur réseau « Immersed
boundary-lattice Boltzmann method »
IBM
Conditions aux limites immergées « Immersed boundary method »
LBGK
Équation de Boltzmann avec le modèle de collision BGK
LBM
Méthode de Boltzmann sur réseau « Lattice Boltzmann method »
MBB
Règle du rebond modifiée « Modified bounce-back »
MRT
Modèle de collision à plusieurs facteur de relaxation « Multi Relaxation Time »
NSERC
« Natural Sciences and Engineering Research Council of Canada »
PIV
Vélocimétrie par imagerie de particules « Particle Image Velocimetry »
RANS
Équations de Navier-Stokes moyennées « Reynolds average Navier-Stokes »
SHI
Sumitomo Heavy Industries
URPEI
Unité de Recherche en Procédés d’Écoulements Industriels
xxii
Lettres latines
Matrice des poids
Paramètre rhéologique de transition du modèle de Carreau-Yasuda
⃗
Vecteur objectif
Espacement entre la pale et le fond de la cuve
Espacement standard entre la pale et le fond de la cuve
Vitesse du son de lattice
Diamètre de l’agitateur
Dimension du problème
Vecteur vitesse
Direction ou vecteur vitesse
Direction d’extrapolation
Erreur par rapport à la solution analytique
Nombre de Froude
Force volumique
et
Populations de particules
et
Populations d’équilibre
Population de non équilibre
Population de post-collision
Fonction d’interpolation
Force volumique
Hauteur de liquide
Tenseur identité
xxiii
Vecteur de quantité de mouvement
Nombre de Knudsen
Constante de puissance
Constante de Meztner-Otto
Numéro de direction¸
Constante de Boltzmann (
)
Dimension caractéristique
Libre parcours moyen d’une particule
Nombre de Mach
Paramètre rhéologique du modèle de puissance
Vitesse de rotation de l’agitateur
Nombre de puissance
Nombre de pompage
Indice de puissance des modèles rhéologiques
Puissance consommée
Pression
Débit caractéristique
Nombre de directions par lattice
Constante des gaz parfaits
Nombre de Reynolds
Nombre de Reynolds généralisé
Rayon
,
Rayons de cylindre
Tenseur du taux de déformation
xxiv
Coordonnées paramétriques
Température
Coordonnée temporelle
Temps post-collision
Temps de mélange
⃗
Vecteur vitesse
⃗
Vitesse de l’obstacle
⃗
Vitesse des points de contrôle
⃗
Vitesse d’une lattice fluide
⃗
Vitesse d’une lattice voisine fluide
⃗
Vitesse extrapolée à une lattice solide
Capacité de la cuve
Coordonnées des points de contrôle
Coordonnées spatiales
Coordonnées d’un point de la frontière
Coordonnées du centre de gravité
Coordonnées d’une lattice fluide
Coordonnées d’une lattice voisine fluide
Coordonnées d’une lattice solide
Lettres grecs
Frontière du domaine
Distance à l’obstacle
Unité de surface
xxv
Opérateur de collision
Paramètre d’extrapolation
̇
Taux de cisaillement
̇
Taux de cisaillement caractéristique
Paramètre d’extrapolation
Fonction de Kronecker
Pas de temps
Pas d’espace ou taille de lattice
Énergie interne
Paramètre de la décomposition de Chapman-Enskog
ou
Viscosité dynamique du fluide
Viscosité à cisaillement nul
Viscosité à cisaillement infini
Temps caractéristique pour le modèle de Carreau-Yasuda
Masse volumique du fluide
Masse volumique moyenne du fluide
Tenseur des contraintes
Facteur de relaxation
Facteur de relaxation de référence à taux de cisaillement nul
⃗
Vecteur de corrections des vitesses
Poids des polynômes Hermitiens
Poids des populations
xxvi
Symboles
Gradient
Dérivée temporelle
Exceptions relatives aux articles
Article 1 :
Diamètre de la cuve
Diamètre de l’agitateur
Domaine de simulation
̇
Tenseur du taux de déformation
Tolérance
Tenseur des contraintes
Vitesse de rotation
Article 2 :
,
Force
Force volumique
Hauteur de liquide
Nombre total de lattices fluides
Diamètre de la cuve
Largeur des contre-pales
̇
Domaine de simulation
Tenseur du taux de déformation
Vitesse de rotation
xxvii
Article 3 :
Hauteur de pale
Hauteur de liquide
Diamètre de la cuve
Largeur des contre-pales
Vitesse de rotation
xxviii
LISTE DES ANNEXES
Annexe A : Intégration de LBGK sur l’espace des vitesses…………………………………….170
Annexe B : La décomposition de Chapman-Enskog……………………………………………172
1
CHAPITRE 1
INTRODUCTION
1.1 Contexte
Ce travail au sein du laboratoire de l’Unité de Recherche en Procédés d’Écoulements Industriels
(URPEI) s’inscrit dans la problématique associée à l’étude des écoulements complexes dans les
cuves agitées. Dans ce laboratoire, les méthodes expérimentales et numériques sont utilisées pour
améliorer la compréhension des phénomènes associés aux écoulements polyphasiques et aux
rhéologies complexes survenant dans les géométries possédant des parties mobiles ou des formes
complexes.
La méthode numérique la plus couramment utilisée à l’URPEI pour la simulation d’écoulements
visqueux est la méthode des éléments finis (FEM). Grâce à cette approche, la caractérisation des
écoulements de nombreux mélangeurs a été réalisée dans les régimes laminaires et transitoires.
Ces résultats ont été obtenus grâce à plusieurs travaux de recherche traitant de la modélisation de
la complexité de la géométrie. Dans la majorité des cas, les simulations se limitent à l’agitation
d’un liquide unique dont les propriétés rhéologiques restent relativement simples.
Cette thèse s’inscrit donc dans l’objectif à long terme de simuler des écoulements polyphasiques
et à rhéologies complexes dans des cuves agitées. L’idée principale est de tirer profit des
particularités de la méthode de Boltzmann sur réseau (LBM) pour faciliter la modélisation de la
complexité de la géométrie et des écoulements. L’utilisation de cette méthode est motivée par les
résultats obtenus par un code LBM, développé à l’URPEI, pour la simulation numérique directe
d’écoulements dans des milieux poreux [1-3]. Ces travaux de recherche ont montré les grandes
capacités de la LBM pour le calcul en parallèle et la facilité à traiter des géométries complexes.
Finalement, il est important de souligner que les conclusions de cette thèse ne se limitent pas à la
simulation des écoulements en cuves agitées, mais sont applicables à tous problèmes comprenant
des écoulements visqueux dans des géométries complexes avec des parties mobiles.
2
1.2 Les écoulements visqueux complexes
Les écoulements visqueux complexes font référence à des écoulements caractérisés par une
prédominance des termes visqueux sur l’inertie, empêchant ainsi la création de turbulence. Les
différentes sources de complexité peuvent se trouver dans la géométrie, la rhéologie du fluide ou
la présence de plusieurs phases. Dans le cadre de cette thèse, dont l’application concerne les
écoulements visqueux en cuves agitées, nous étudierons principalement la problématique du
traitement de la complexité de la géométrie. Dans la dernière partie, l’impact de la rhéologie sera
aussi étudié.
1.2.1 Les écoulements dans des géométries complexes
L’écoulement dans un milieu poreux est un exemple très connu d’écoulement visqueux
complexe. Comme illustré dans la Figure 1.1, la surface de contact entre le fluide et le solide est
très grande et de forme complexe.
Figure 1.1: Espace poreux d'un morceau de grès [4].
Avec les méthodes standards de CFD, la discrétisation de la partie fluide est très difficile à
réaliser et génère un très grand nombre d’éléments (ou de cellules), ce qui complique le calcul de
la solution. Afin de contourner cette difficulté, il est possible de modéliser l’action du milieu
poreux sur le fluide par un modèle de Darcy [5]. L’introduction d’un modèle supplémentaire
entraine une réduction de la précision des simulations, mais permet d’obtenir un résultat dans un
temps de calcul raisonnable (voir un exemple d’applications dans [6]). À l’inverse, l’utilisation de
la LBM facilite grandement la simulation numérique directe de l’écoulement dans un milieu
poreux. En effet, à l’aide d’un maillage structuré cubique et d’une approximation des frontières
3
par faces, l’étape de discrétisation devient possible [1]. Comme la méthode est explicite et très
efficace pour le calcul parallèle, elle permet d’utiliser un très grand nombre d’éléments (lattices)
[3]. Par exemple, l’utilisation de la LBM a permis la simulation numérique directe de
l’écoulement dans un pot catalytique à l’aide de
lattices fluides [7], comme illustré
dans la Figure 1.2.
Figure 1.2: Simulation LBM dans un pot catalytique (tirée de [7])
Les géométries avec parties mobiles ajoutent une degré supplémentaire de difficulté puisqu’il est
nécessaire d’adapter la discrétisation en fonction du mouvement des objets. Comme la
reconstruction du maillage dans le temps est une tâche très difficile et très longue, ce sont les
méthodes de maillages glissants [8] ou de conditions aux limites immergées [9, 10] qui sont
préférées. Ce point sera traité plus en détail dans la section 1.3.3 pour le cas des cuves agitées.
1.2.2 Les modèles rhéologiques
L’écoulement d’un fluide dans une géométrie donnée est très largement affecté par sa rhéologie,
il est donc important de modéliser précisément ses effets. Comme dans le cas général la valeur de
la viscosité d’un fluide dépend d’un très grand nombre de paramètres, le but de la rhéologie est
de caractériser les dépendances les plus significatives afin de modéliser le comportement des
4
fluides. Dans le cadre de cette thèse, l’utilisation des modèles sera restreinte aux comportements
newtoniens et pseudo-plastiques.
Dans le cas des fluides newtoniens et non newtoniens sans effet d’élasticité, la relation entre le
tenseur des contraintes
et le tenseur du taux de déformation
est définie comme :
(1.1)
avec
la pression,
la viscosité dynamique, le tenseur identité et
Dans le cas newtonien, la viscosité est constante (
fonction du taux de cisaillement
̇
√
⃗
⃗
.
). Dans le cas non newtonien, elle est
. Les comportements les plus courants sont
représentés dans la Figure 1.3(a). Afin de décrire un comportement pseudo-plastique (ou
rhéofluidifiant), le modèle le plus simple est la loi de puissance [11] :
̇
où les paramètres
et
(1.2)
sont déterminés empiriquement.
Afin de corriger le modèle aux valeurs limites du taux de cisaillement (
), il est possible d’utiliser le modèle de Carreau-Yasuda [11] :
| ̇|
[
avec
et
| ̇|
et
̇
]
les viscosités à taux de cisaillement nul et infini,
(1.3)
le temps caractéristique et
un
paramètre qui décrit la transition entre le comportement newtonien et la loi de puissance. Ce
modèle permet d’inclure deux plateaux newtoniens aux valeurs limites du taux de cisaillement,
comme illustré dans la Figure 1.3(b).
Dans le cas général, l’intégration d’un modèle rhéologique dans les équations de Navier-Stokes
rend leur résolution plus difficile [12]. Par exemple, il a été montré que lors de l’utilisation d’un
modèle de loi de puissance, il est très difficile de simuler un liquide avec un indice de puissance
[13]. Dans le cas de la LBM, l’intégration d’un modèle rhéologique sans effet
d’élasticité peut se faire d’une manière naturelle qui ne perturbe pas les propriétés du schéma
5
[14]. Par exemple, on notera des simulations réalisées avec des modèles de loi de puissance [15,
16] et de Carreau-Yasuda [17, 18]. Ce point sera traité plus en détail dans la section 2.5.
Figure 1.3: Illustrations des comportements rhéologiques
1.2.3 Les écoulements polyphasiques
Dans le cadre de cette thèse, les écoulements polyphasiques ne seront pas étudiés. Cependant, il
est intéressant d’aborder cette problématique puisqu’elle s’inscrit dans la continuité de ce travail.
Afin de réaliser la simulation directe d’un écoulement polyphasique, les problématiques
soulevées précédemment reviennent. En effet, pour décrire adéquatement l’interaction entre les
phases, il est nécessaire de discrétiser précisément la frontière qui les sépare, comme illustré dans
la Figure 1.4 dans le cas d’une suspension. De la même façon, la rhéologie des fluides en
présence va affecter largement l’écoulement global.
Figure 1.4: Discrétisation d'une suspension (simulation 2D) [19]
6
Comme il n’est généralement pas réalisable d’effectuer une simulation numérique directe (DNS)
d’un écoulement polyphasique à l’échelle industrielle, il est nécessaire de modéliser certains
phénomènes physiques au sein de l’écoulement. Le choix de la méthode et des modèles utilisés
est un compromis entre la taille de la discrétisation et l’échelle physique étudiée. Le lecteur
intéressé pourra se référer à [20] pour un exemple de stratégie multi-échelle appliquée à la
simulation de lits fluidisés gaz-liquide. En fonction de l’échelle physique désirée, il est possible
d’utiliser des méthodes continues monophasiques (ou 1-fluide) [21, 22], dont les propriétés du
fluide modèle doivent caractériser l’écoulement polyphasique au complet, ou des méthodes
continues avec 2 fluides [23-25] qui permettent de modéliser l’écoulement polyphasique comme
l’enchevêtrement de deux fluides. Il est à noter que l’augmentation des capacités de calcul a
permis l’application de stratégies de simulation à l’échelle intermédiaire, comme celle des
particules discrètes (DPM) [26-28] dans le cas des suspensions. Avec cette approche, il y a un
couplage entre les particules et le liquide, c’est-à-dire que le liquide agit sur les particules et que
celles-ci agissent aussi sur le liquide. Les trajectoires des particules sont déterminées grâce à la
seconde loi de Newton (approche Lagrangienne) et les collisions sont modélisées (voir [29, 30]
pour un modèle de type ressort-amortisseur). Contrairement à la DNS, la phase liquide est
continue dans tout le domaine (pas de frontière avec les particules), ce qui permet de réduire le
nombre d’éléments utilisés et donc les temps de calcul.
1.3 Le mélange visqueux en cuves agitées
Les mélangeurs sont couramment utilisés dans l’industrie dans le but de disperser une substance,
de mélanger plusieurs phases ou de conserver l’homogénéité d’une mixture. Les mélangeurs
industriels présentent une large variété de formes et de fonctionnements afin de s’adapter à toutes
les situations. Que les systèmes de mélange soient dynamiques ou statiques, qu’ils utilisent des
pales raclantes ou des hélices, ils profitent de l’écoulement en leur sein pour réaliser ou conserver
le mélange de produits.
D’une manière générale, la façon la plus rapide et la plus simple de mélanger est d’utiliser la
turbulence. En effet, grâce à l’action des tourbillons présents aux différentes échelles,
l’enchevêtrement des différents produits est rapidement réalisé. À l’inverse, dans le cas des
liquides visqueux, la turbulence ne peut être utilisée car elle entrainerait une consommation
d’énergie trop élevée. Le mélange visqueux est expliqué plus en détails dans la section suivante.
7
1.3.1 Les géométries des mélangeurs
Les mélangeurs spécialisés pour les liquides visqueux compensent l’impossibilité d’utiliser la
turbulence par des géométries complexes. Par exemple, la Figure 1.5 présente 3 types de
mélangeurs : un mélangeur statique, un mélangeur planétaire et un mélangeur raclant.
(a)
(b)
(c)
Figure 1.5 : Exemples de mélangeurs spécialisés pour les liquides visqueux : (a) mélangeur
statique Kenics, (b) mélangeur planétaire tripale et (c) double ruban hélicoïdal.
Dans ces trois exemples, la géométrie des obstacles ou des parties mobiles a pour but de plier et
d’étirer les lignes de courant afin de générer le mélange. Autrement dit, l’efficacité d’un
mélangeur spécialisé pour les liquides visqueux se caractérise par sa capacité à créer du
cisaillement (ou de l’élongation). D’un point de vue industriel, il est important que ce
cisaillement soit réalisé à moindre coût. La notion du coût de mélange est généralement abordée à
partir de la consommation de puissance du système de mélange.
Dans cette thèse, ce sont les mélangeurs discontinus « batch mixer » spécialisés pour les liquides
visqueux qui sont étudiés. Ils sont caractérisés par la présence d’une cuve et d’au moins un
agitateur, comme présenté dans la Figure 1.6, et sont très largement utilisés pour les procédés de
fabrication qui nécessitent plusieurs phases. Pouvant être associés à un système de contrôle de la
température, ils permettent aussi bien de contrôler l’homogénéité d’une mixture durant toute la
durée du procédé que de disperser un additif. Ces systèmes doivent être capables de prévenir
toutes stagnations du produit ou toutes créations de zones mortes. Pour ce faire, les pales doivent
racler le bord et le fond de la cuve et être capables de générer une circulation axiale. Ultimement,
un système de mélange spécialisé pour les fluides visqueux doit être capable de générer un fort
8
cisaillement (mélange dispersif) et une circulation axiale dans tout le volume de la cuve (mélange
distributif).
Figure 1.6: Mélangeur Maxblend (SHI, 2008)
L’efficacité du mélange est aussi largement affectée par les propriétés physiques de la mixture.
Par exemple, un système très efficace pour des liquides newtoniens peut voir ses performances
diminuer largement pour des rhéologies complexes. D’une manière générale, le choix d’un
système de mélange est basé sur la situation la plus problématique, c’est-à-dire la phase du
procédé où la viscosité est la plus élevée ou lorsque que la rhéologie est la plus complexe.
1.3.2 Les grandeurs caractéristiques des mélangeurs
Afin de comparer les systèmes de mélange entre eux, trois critères peuvent être utilisés : le temps
de mélange, la consommation de puissance et le pompage. L’utilisation de ces grandeurs reste
limitée car elles dépendent du régime de l’écoulement, de la rhéologie et des paramètres
géométriques.
Dans le cas d’un système de mélange avec une seule pale, le nombre de Reynolds est défini
comme :
(1.4)
avec
la masse volumique du fluide,
la vitesse de rotation et
le diamètre de l’agitateur.
9
Cette définition peut être généralisée aux cas non newtoniens. Dans le cas d’un liquide pseudoplastique, le concept de Metzner et Otto [31] propose de définir la viscosité caractéristique à
l’aide du taux de cisaillement ̇
,
étant une constante qui dépend seulement de la
géométrie du système. Le nombre de Reynolds généralisé est donc :
(1.5)
Le temps de mélange est une notion prépondérante pour la comparaison des mélangeurs car il
permet de caractériser le temps nécessaire pour obtenir l’homogénéité du mélange. Cette
grandeur peut être mesurée expérimentalement grâce à des méthodes de décoloration [32, 33]. Le
principe est basé sur l’injection d’un réactif qui modifie les propriétés visuelles du mélange.
Ainsi, pour chaque nombre de Reynolds étudié, le temps de mélange
est atteint quand le
réactif a agi sur le volume complet du fluide. De cette manière, il est possible de construire des
courbes
versus
afin de comparer les systèmes de mélange. Par exemple, la Figure 1.7
montre que les rubans hélicoïdaux présentent un temps de mélange faible pour des régimes très
laminaires (
), à l’inverse ce sont les turbines qui sont les plus performantes à hauts
nombres de Reynolds.
Figure 1.7: Courbes de temps de mélange pour différents systèmes d'agitation (données
provenant de Yamato et al., 1998 [34])
10
Le coût du mélange est généralement caractérisé à travers la consommation de puissance
système d’agitation. Pour ce faire, le nombre adimensionnel de puissance
du
est défini comme :
(1.6)
Ce nombre est constant dans le régime turbulent, comme illustré dans la Figure 1.8 dans le cas du
mélangeur Maxblend. À l’inverse, en régime laminaire, c’est la constante adimensionnelle de
puissance
qui est constante. On remarque que grâce au concept de concept de
Metzner et Otto introduit à la page précédente, il est possible de superposer les courbes de
puissance newtonienne et non newtonienne.
Figure 1.8: Courbes de puissance pour le mélangeur Maxblend (données provenant de
Guntzburger et al., 2012 [35]).
La création d’une circulation axiale dans le système de mélange est cruciale pour l’homogénéité
de la mixture. Afin de mesurer la capacité de pompage, le nombre de pompage
est défini
comme :
(1.7)
11
La valeur du débit caractéristique
est généralement approximée par l’intégration du champ de
vitesse passant au travers d’un groupe de plans [13, 36]. La répartition et la forme de ces plans
dépendent de la géométrie du système de mélange.
Les valeurs des grandeurs caractéristiques associées au temps de mélange et au pompage peuvent
varier avec la rhéologie et donc caractériser un gain ou une perte de performance du système de
mélange. Afin d’obtenir les valeurs de
,
et
, il est possible d’utiliser les approches
expérimentales et numériques. Chacune possédant des avantages et des inconvénients, il est
intéressant d’utiliser les deux approches conjointement afin de maximiser l’information sur un
système de mélange.
1.3.3 La simulation du mélange visqueux
L’utilisation des méthodes numériques permet de compléter les informations obtenues grâce aux
approches expérimentales. Par exemple, elles permettent facilement de visualiser des
écoulements tridimensionnels globaux ou, à l’inverse, d’obtenir des informations locales sans
modifier l’écoulement en lui-même. Pour autant, les méthodes numériques possèdent aussi leurs
propres limites. Dans le cas du mélange visqueux, les difficultés apparaissent dans le traitement
de la géométrie, de la rhéologie et des écoulements polyphasiques.
Dans le cas des mélangeurs discontinus, l’écoulement est généré par le mouvement des pales. Il
est donc primordial que la méthode numérique utilisée soit capable de modéliser précisément
l’impact de la pale sur le liquide environnent. Cette tâche est non triviale car la géométrie de la
pale est complexe et sa position varie dans le temps. Aussi, il n’est pas réalisable de remailler la
zone liquide à chaque itération car cela entrainerait des temps de calcul trop importants. Afin de
contourner cette difficulté, plusieurs solutions sont disponibles dans la littérature. Premièrement,
lorsque que la forme de la géométrie n’évolue pas dans le temps, il est possible d’appliquer un
changement de repère, comme présenté dans la Figure 1.9(a). Ainsi, à l’aide d’un repère
Lagrangien associé à la pale [37], il est possible de résoudre les équations de Navier-Stokes sur
un maillage fixe. Il suffit simplement de corriger les équations en incluant les forces de Coriolis
et centrifuge. Malheureusement, cette approche est très limitative car elle ne permet que la
simulation d’une pale centrée dans une cuve sans contre-pale. Afin de dépasser cette limitation, il
est possible de décomposer le volume en deux parties cylindriques coaxiales (voir Figure 1.9(b)).
La première partie, associée au repère naturel, comprend la cuve et les contre-pales. La deuxième
12
partie, associée à un repère Lagrangien, n’inclut que la pale. De cette manière, il est possible de
résoudre les équations de Navier-Stokes dans deux géométries fixes, la difficulté résidant ici dans
la définition de la communication entre les zones. Avec à la méthode dite des repères multiples «
mutiple reference frame » (voir Brucato et al. [38] pour un exemple d’application), il est possible
d’obtenir une approximation stationnaire de l’écoulement. Pour pouvoir obtenir l’évolution de
l’écoulement dans le temps, il faut utiliser la méthode du maillage tournant « sliding mesh » [8],
qui permet de prendre en compte la rotation de la pale. Cependant, comme il est difficile de gérer
les communications entre les deux zones, la précision peut être dégradée [39]. Ces techniques
sont difficilement applicables aux agitateurs raclants car elles nécessitent un écart important entre
la pale et la cuve. De plus, le mouvement de la pale est limité à une rotation autour de son axe.
Afin de dépasser ces limitations, il est possible d’utiliser une méthode de type conditions aux
limites immergées « immersed boundary method (IBM) » [9, 10, 40]. Le principe est de
considérer un domaine entièrement fluide à l’intérieur duquel, à différents points de contrôle, la
vitesse du fluide est contrainte. Ainsi, comme présenté dans Figure 1.9(c), à l’aide d’un groupe de
points décrivant la surface de la pale, le fluide environnant est contraint à suivre le mouvement de
l’agitateur. Dans ce cas, le traitement de la variation de la géométrie se limite à la simple
modification des coordonnées des points de contrôle. Cette méthode est plus générique mais la
précision du champ de vitesse autour des points de contrôle n’est pas toujours maitrisée.
Figure 1.9: Illustrations des différentes approches de gestion de la géométrie. (a) changement de
repère, (b) utilisation de deux repères et (c) conditions aux limites immergées.
aux repères fixes par rapport à la cuve et à la pale, respectivement.
et
réfèrent
13
Grâce à ces différentes approches, des nombreux systèmes de mélanges ont été étudiés. Il ressort
de la littérature que c’est principalement la méthode des éléments finis « finite element method
(FEM) » qui est utilisée, le plus généralement pour des régimes laminaire et transitoire. Ainsi,
grâce à l’utilisation d’un repère Lagrangien, les hydrodynamiques d’un mélangeur Paravisc®
[41], d’un malaxeur à pâte [42], de rubans hélicoïdaux (DHR) [12, 37, 43] et d’un mélangeur
Maxblend® [13, 36, 43, 44] ont été étudiées avec succès. Pour des géométries plus complexes,
les méthodes de type conditions aux limites immergées ont permis de simuler les écoulements
dans des mélangeurs planétaires [44, 45] et dans un mélangeur Maxblend avec contre-pales [13,
36, 44]. On notera aussi le couplage d’une méthode de conditions aux limites immergées avec un
maillage glissant pour simuler l’hydrodynamique du Superblend® [46], constitué d’un Maxblend
et d’un DHR. Pour ce qui est de la rhéologie, la FEM a permis de simuler des comportements
pseudo-plastiques avec des modèles de type loi de puissance [13, 41] et Carreau-Yasuda [42],
l’indice de puissance le plus petit étant
. On notera aussi l’utilisation d’un modèle
simplifié de comportement viscoélastique [12] et d’un fluide à seuil de type Bingham [47].
Pour la méthode des volumes finis « finite volume method (FVM) », son utilisation dans le cadre
du mélange est généralement couplée à la méthode de maillage glissant et se situe dans le régime
turbulent. En effet, la plupart des articles sont basés sur l’utilisation du logiciel Fluent®, lequel
dispose d’une méthode de type maillage glissant. Ce sont généralement des turbines Rushton qui
ont été étudiées [48-51] dans le cas de régimes d’écoulements turbulents (donc faiblement
visqueux). Dans le cas d’écoulements à nombres de Reynolds plus faibles, on notera l’étude d’un
mélangeur coaxial constitué d’une ancre et d’une turbine [52].
À partir de ces travaux, il apparait que la simulation de l’hydrodynamique au sein des mélangeurs
industriels pour liquides visqueux est principalement faite à l’aide de la méthode des éléments
finis. Les simulations sont généralement monophasiques (agitation d’un liquide) et se limitent à
des rhéologies simples (
). Ces restrictions s’expliquent par le très grand nombre
d’éléments (et donc d’équations) nécessaires pour décrire la complexité de la géométrie. À cela
s’ajoute l’intégration de la rhéologie qui, à cause de la non linéarité inhérente à ces modèles,
entraine une dégradation des propriétés du système d’équations à résoudre. Ces difficultés
mènent généralement à un compromis entre précision et temps de calcul, limitant alors la
simulation de problèmes plus complexes, comme le mélange de mixtures polyphasiques ou à
14
rhéologies complexes. Afin de contourner ces difficultés, il est pertinent de tester de nouvelles
approches numériques.
Comme la méthode de Boltzmann sur réseau « lattice Boltzmann method (LBM) » est très
efficace en parallèle [2], qu’elle permet de traiter des géométries complexes [3, 53] et d’intégrer
facilement des modèles rhéologiques [16, 27, 54], nous proposons de l’utiliser dans le cas des
écoulements visqueux en cuves agitées.
15
1.4 Objectif principal de la thèse
Au regard des difficultés mentionnées précédemment pour la simulation du mélange visqueux, il
est maintenant possible de formuler l’objectif principal de cette thèse :
Développer une méthode de Boltzmann sur réseau efficace pour la simulation
d’écoulements visqueux en cuves agitées.
L’hypothèse associée à cet objectif est que l’utilisation de la LBM va permettre d’accroitre les
possibilités de simulations au travers de la prise en compte efficace de la géométrie et de la
rhéologie.
16
CHAPITRE 2
REVUE DE LITTÉRATURE : MÉTHODE DE
BOLTZMANN SUR RÉSEAU
Dans ce chapitre, les aspects fondamentaux de la méthode de Boltzmann sur réseau sont
brièvement introduits afin de faire ressortir les notions importantes pour la simulation du mélange
visqueux. Le lecteur intéressé pourra se référer aux thèses de Kandhai [55], Latt [56], Malaspinas
[14] et Matilas [57] et au livre de Succi [58]. Il est important de mentionner que le
développement de la LBM peut être obtenu de différentes manières. En effet, la méthode peut
être vue comme une version continue des automates cellulaires ou une version discrétisée de
l’équation de Boltzmann. De plus, dans chacune de ces deux approches, il existe plusieurs
variantes et les étapes peuvent être effectuées dans un ordre différent.
Sans perte de généralité, il ne sera présenté dans la prochaine section qu’un seul développement,
principalement basé sur Malaspinas [14] et Shan [59]. De cette manière, il sera possible de faire
apparaitre les paramètres numériques et d’introduire les différentes hypothèses associées à la
LBM. Il est important de souligner que le développement de Boltzmann à Navier-Stokes se
limitera à un fluide athermal faiblement compressible et que les forces volumiques ne sont pas
considérées.
2.1 De Boltzmann à Navier-Stokes
L’idée sous-jacente à la méthode de Boltzmann sur réseau (LBM) est que le comportement
macroscopique d’un fluide est indépendant de la nature des particules qui le composent. Ainsi, en
définissant des particules fictives dont le comportement peut être facilement simulé, il est
possible d’obtenir une approximation de l’écoulement du fluide (voir la citation de Feynman dans
[55]).
2.1.1 Équation de Boltzmann sous forme continue
Contrairement aux équations de Navier-Stokes, l’équation de Boltzmann est basée sur une échelle
physique différente, l’échelle mésoscopique, qui permet d’étudier le comportement d’un groupe
de particules (théorie cinétique des gaz). Si l’on introduit
de présence de particules (ou population) au point
comme étant une probabilité
, avec une vitesse
évolution peut être déterminée par l’équation de Boltzmann :
et au temps , son
17
(2.1)
avec
l’opérateur de collision interparticulaire.
Grâce à ces populations, il est possible de calculer les grandeurs macroscopiques (masse
volumique, quantité de mouvement et énergie interne) par intégration dans l’espace des vitesses :
∫
⃗
(2.2)
(2.3)
∫
∫
(2.4)
⃗
Ces grandeurs sont généralement appelées ‘moments’ et correspondent chacune à un ordre
différent en vitesse .
Dans le cas d’un gaz monoatomique, on notera que l’énergie interne peut aussi s’exprimer
comme
température et
avec
√
la dimension spatiale du problème et
la constante des gaz parfaits (
où
est la
).
Comme l‘équation (2.1) est difficile à résoudre à cause de la non linéarité de l’opérateur de
collision, Bhatnagar, Gross et Krook (BGK) [60] ont proposé un opérateur de collision simplifié,
modélisé par une relaxation vers l’équilibre. L’hypothèse sous-jacente est que la collision
entraine les populations vers un état d’équilibre. De cette manière, l’équation (2.1) prend la forme
suivante (appelée LBGK) :
(
avec le facteur de relaxation et
)
(2.5)
la population d’équilibre. Cette dernière est définie par une
fonction de distribution de Maxwell-Boltzmann :
(
⃗
)
(2.6)
18
Par définition, les moments peuvent être indépendamment calculés avec les populations ou les
populations d’équilibre :
⃗
∫
∫
∫
∫
∫
⃗
(2.7)
(2.8)
∫
(2.9)
⃗
En intégrant l’équation LBGK (2.5) sur l’espace des vitesses (voir Annexe A), il est possible d’en
déduire les équations de conservations de la masse et de la quantité de mouvement :
⃗
⃗
avec ⃗ ⃗
⃗
(2.10)
⃗⃗
(2.11)
⃗ , le produit tensoriel du vecteur ⃗ avec lui-même, et
le tenseur des contraintes
calculé comme :
∫
(
)
c'est-à-dire que le tenseur est exprimé comme un moment d’ordre 2 en vitesse (
(2.12)
).
Si cette définition du tenseur assure sa symétrie, elle ne permet pas de fermer le système
d’équations (2.10) et (2.11). En effet, il comporte 4 équations pour 10 inconnues (1 masse
volumique, 3 vitesses et 6 contraintes).
2.1.2 Équation de fermeture
Afin de fermer le système d’équations (2.10) et (2.11), obtenu à partir de l’équation de
Boltzmann, il est nécessaire d’exprimer le tenseur des contraintes à partir des autres grandeurs
macroscopiques.
Tout d’abord, il faut appliquer la décomposition de Chapman-Enskog [61, 62] à l’équation
LBGK (2.5). Cette décomposition est basée sur une séparation des phénomènes en fonction de
19
leur échelle de temps (l’advection est plus rapide que la diffusion). Dans cette procédure (voir
Annexe B), les populations et les dérivées sont décomposées en plusieurs termes relatifs à chaque
échelle de temps. Par exemple, les populations sont exprimées sous la forme suivante :
(2.13)
Il est important de noter que le paramètre
nombre de Knudsen
avec
est généralement identifié comme étant le
le libre parcours moyen d’une molécule et
une
dimension caractéristique de l’écoulement. Cette étape ajoute donc à la LBM l’hypothèse
d’écoulement à faible
.
Lorsque la décomposition est appliquée à l’équation LBGK (2.5), il est possible de regrouper les
termes par puissance de :
L’équation (2.14) montre que la population
Ordre 0
(2.14)
Ordre 1
(2.15)
correspond à la fonction d’équilibre. Cela permet
de montrer que la décomposition des populations dans l’équation (2.13) revient à séparer les
populations entre un terme d’équilibre
et de non équilibre
. Cette décomposition sera
utilisée ultérieurement dans le document.
Afin de déterminer l’équation de fermeture, l’équation (2.15) est projetée sur une base
Hermitienne. L’intérêt de l’utilisation de cette base est double puisque les coefficients de
l’expansion des populations peuvent s’exprimer à partir des moments et qu’il existe une relation
de récurrence entre les coefficients. Dans le cas d’un fluide athermal faiblement compressible,
une troncature à l’ordre 2 de la projection des populations est suffisante pour conserver la masse
et la quantité de mouvement. De cette manière, une relation entre les coefficients peut être
établie, ce qui permet de relier les grandeurs macroscopiques entre elles. L’équation de fermeture
obtenue est alors :
(2.16)
20
avec
la pression et
où
, avec
la constante de Boltzmann et
la masse
d’une particule.
On notera que l’obtention cette équation nécessite l’introduction de la loi des gaz parfaits sous la
forme :
(2.17)
La troncature sur
à l’ordre 2 permet de réécrire la fonction d’équilibre
(
avec
⃗
⃗
⃗ ⃗
sous la forme :
)
(2.18)
le poids associé à .
Cette définition n’est valable que pour un fluide athermal faiblement compressible et ajoute
l’hypothèse de petits nombres de Mach (
|⃗ |
).
2.1.3 Discrétisation en vitesse
L’étape de discrétisation en vitesse est basée sur le principe que les intégrales associées au calcul
des coefficients de l’expansion des populations peuvent être exactement évaluées avec une
somme pondérée, basée sur une quadrature de Gauss-Hermite. Le choix de cette quadrature est
justifié par le fait que le domaine d’intégration est ]
[. Comme les intégrales sont dans
l’espace des vitesses, l’application de la quadrature de Gauss-Hermite fait apparaitre des points
de Gauss correspondant à des vitesses discrètes
, appelées aussi directions.
Le choix de l’ordre de la quadrature dépend de la troncature sur la population
faite dans la
section précédente. Comme précisé auparavant, nous nous sommes limités aux équations de
Navier-Stokes pour un fluide athermal faiblement compressible. Dans ce cas, la population
a
été tronquée à l’ordre 2, la quadrature de Gauss-Hermite s’applique donc à des polynômes dont
l’ordre est au maximum 4 (voir Malaspinas [14] et Shan [59]).
Les moments peuvent ainsi être calculés exactement à partir d’un nombre fini
. En introduisant la population
de poids
de populations
, associée à la direction
, comme :
21
(2.19)
les moments discrétisés deviennent
(2.20)
∑
(2.21)
⃗
∑
et la fonction d’équilibre devient :
(
où
⃗
⃗
⃗
)
(2.22)
√ , avec la vitesse de propagation qui sera explicitée dans la section suivante.
De même, dans le tenseur des contraintes, l’expression utilisée pour décrire la viscosité devient :
(2.23)
D’un point de vue pratique, la discrétisation de l’espace des vitesses en un ensemble de directions
{ }
permet de définir les lattices. Cette étape est un point clé de la méthode
LBM car elle permet de façonner leur forme, c’est-à-dire la structure de communication de la
méthode. Les lattices obtenues sont nommées DdQq, avec d la dimension du problème et q le
nombre total de directions. Dans le reste de cette thèse, c’est la lattice D3Q15 [63] qui sera
utilisée. En effet, d’après nos tests, lorsqu’elle est comparée aux lattices D3Q19 et D3Q27, elle
permet de réduire les temps de calculs tout en conservant une précision similaire. Il est à noter
que l’utilisation de la lattice D3Q15 peut générer des effets en damier sur les champs de pression
et de vitesse [64]. Si ces effets sont négligeables pour des nombres de Reynolds modérés, il est
conseillé d’utiliser les lattices D3Q19 ou D3Q27 pour les écoulements turbulents. L’ensemble
associé { } est défini dans le Tableau 2.1 et est illustré dans la Figure 2.1.
22
Grâce à cette discrétisation en vitesse, il est maintenant possible d’exprimer l’équation LBGK
(2.5) sous la forme :
(
Tableau 2.1: Directions
0
1
2
0
et coefficients
pour la lattice D3Q15.
3
4
5
6
0
0
0
0
0
0
1/9
1/9
0
0
0
0
0
0
0
0
2/9
1/9
1/9
1/9
1/9
(2.24)
)
7
8
9
10
11
12
13
1/72 1/72 1/72 1/72 1/72 1/72 1/72
Figure 2.1: Représentation de la lattice D3Q15
2.1.4 Discrétisation spatiale et temporelle
Après avoir effectué la discrétisation de l’espace des vitesses, il faut maintenant discrétiser
temporellement l’équation (2.24), c’est-à-dire que l’on introduit le pas de temps
. On notera
que dans le schéma LBM, les populations doivent se déplacer exactement d’une lattice.
Autrement dit,
devient :
, ce qui définit la discrétisation spatiale. Dans ce cas, le calcul de
23
(2.25)
√
L’intégration directe de l’équation (2.24) donne :
(2.26)
∫
Si on applique la méthode des trapèzes pour calculer l’intégrale, on obtient :
(
(2.27)
)
En introduisant le changement de variable suivant :
̃
(
(2.28)
)
l’équation (2.27) devient
̃
̃
̃
(̃
)
(2.29)
avec
̃
(2.30)
Grâce à ce changement de variable, on obtient une formulation explicite d’ordre 2 en temps. Il est
important de noter que le changement de variable affecte la définition de la viscosité. L’équation
(2.23) devient :
24
( ̃
(2.31)
)
Par contre, les calculs de la masse volumique et de la vitesse ne sont pas affectés. Dans le cas
plus général où une force volumique est considérée, le changement de variable entraine
l’introduction de cette force dans le calcul des vitesses [14, 65]. Dans le reste de la thèse, le tilde
(~) sera omis par mesure de simplicité et seules les populations avec changement de variable
seront considérées.
2.2 Équation de collision-propagation
Commençons par résumer le schéma LBM présenté dans la section 2.1. L’équation LBGK
discrétisée prend la forme :
(
)
(2.32)
avec la fonction d’équilibre définie comme :
⃗
(
où les poids
⃗
dépendent de la lattice utilisée et la vitesse du son
⃗
)
(2.33)
est définie comme :
(2.34)
√
Dans chaque lattice, une masse volumique et un vecteur vitesse peuvent être calculés directement
à partir des populations :
∑
(2.35)
(2.36)
⃗
∑
25
Le paramètre de relaxation est relié à la viscosité par la relation suivante :
(2.37)
Dans la section 2.1, nous avons montré que la LBM permet d’obtenir une solution des équations
de Navier-Stokes incompressibles et athermales, sous contrainte de faible nombre de Mach et de
Knudsen. Dans le cadre de la LBM, ces deux nombres s’expriment comme :
|⃗ |
(2.38)
(2.39)
avec
une dimension caractéristique du problème étudié.
Afin de déterminer l’ordre de convergence spatiale, il est nécessaire d’appliquer une analyse
asymptotique de l’équation de lattice Boltzmann [66, 67]. Ainsi, en utilisant l’échelle diffusive
(i.e.
), on montre que la LBM permet d’obtenir une solution des équations de Navier-
Stokes (vitesse et pression) avec une précision d’ordre 2 en espace sur la vitesse. L’utilisation de
l’échelle diffusive (
) entraine une précision en temps d’ordre 1. D’un point de vue
pratique, lorsque l’on cherche à vérifier l’ordre de convergence spatiale, l’utilisation de l’échelle
diffusive implique de conserver une valeur fixe du facteur de relaxation , ce qui permet de
garder un ratio
constant. Il est important de souligner que les développements
mathématiques associés à l’étude de la convergence se limitent généralement à des domaines
périodiques. Pour d’autres configurations, comme des conditions aux limites de Dirichlet en
vitesse, la convergence spatiale est largement affectée [67].
Il est aussi intéressant de regarder plus précisément les calculs associés à la LBM. Il est possible
de réécrire l’équation (2.32) sous la forme suivante :
(
)
(2.40)
(2.41)
26
l'équation (2.40) correspond à l’étape de collision et l’équation (2.41) à celle de propagation. On
remarque alors que les calculs nécessaires à l’application du schéma LBM, c’est-à-dire les
équations (2.40), (2.33), (2.35) et (2.36), sont simples et locaux. La localité signifie que les
calculs associés à une lattice ne nécessitent pas d’autres informations que celles présentes dans la
lattice. La seule étape nécessitant un transfert de l’information étant l’étape de propagation,
définie par l’équation (2.41). On remarque alors que le transfert d’informations est structuré par
les directions
. Toutes ces propriétés donnent à la LBM une très bonne efficacité pour le calcul
parallèle. Il est important de souligner que ces avantages ne sont pas tous intrinsèquement reliés à
la LBM, mais pourraient s’appliquer à d’autres schémas utilisant un maillage structuré cartésien.
Les avantages intrinsèques de la LBM sont plutôt reliés à l’intégration simple de modèles de
phénomènes physiques dont l’origine est microscopique. Par exemple, on peut citer l’intégration
des modèles rhéologiques [16, 54, 55] ou la modélisation de l’interaction entre les phases pour
les écoulements polyphasiques [68-70]. L’intégration de la rhéologie sera traitée plus en détail
dans la section 2.5.
2.3 Définition des conditions aux limites
Les conditions aux limites servent à contraindre la solution numérique sur les frontières du
domaine de simulation. En effet, pour pouvoir simuler un écoulement donné, les conditions aux
limites doivent modéliser les phénomènes physiques qui apparaissent aux différentes surfaces du
volume occupé par le fluide. Les exemples les plus courants de tels phénomènes sont les surfaces
libres, l’attachement des fluides à une paroi solide et les profils d’entrée et de sortie dans les
canaux. Puisque l’échelle physique des écoulements simulés est macroscopique, il est nécessaire
d’imposer des conditions aux limites macroscopiques sur les frontières du domaine. Comme nous
l’avons vu dans la section 2.1, contrairement aux méthodes qui résolvent directement les équation
de Navier-Stokes, la LBM utilise une variable primaire mésoscopique (une population de
particules
), ce qui complique largement l’imposition d’une contrainte macroscopique. En effet,
s’il est très simple de calculer la masse volumique et la vitesse à partir des populations, comme
défini dans les équations (2.35) et (2.36), le calcul inverse n’a pas de solution unique. Ceci
explique pourquoi les conditions aux limites sont une difficulté intrinsèque à la LBM et
requièrent un changement de paradigme par rapport aux méthodes classiques de la CFD. Le but
27
de cette section est de donner un aperçu des stratégies d’imposition de vitesses non nulles sur des
frontières courbes, des explications plus détaillées se trouvant dans l’article 1 (section 4.2).
2.3.1 La règle du rebond
Afin d’imposer une vitesse nulle sur une frontière, l’approche la plus couramment utilisée est
celle de la règle du rebond « bounce-back rule » [71, 72], qui permet de définir les populations
manquantes comme :
(2.42)
avec
la direction opposée à
et
la population post-collision définie par l’équation (2.40).
Figure 2.2: Illustration de la règle du rebond (tirée de l'article 1)
Comme illustrée dans la Figure 2.2, la population
qui devrait être perdue (car propagée
vers une zone solide ou en dehors du domaine) est utilisée pour calculer la direction opposée.
Cette stratégie peut être étendue à l’imposition de vitesse non nulles [72] de la manière suivante :
⃗
(2.43)
avec ⃗ la vitesse de l’obstacle à la frontière.
Il est important de souligner que la position de la frontière est toujours considérée comme étant
située à
, c’est-à-dire au bord de la lattice fluide. Lorsqu’appliquée à mur droit
positionné exactement aux bords des lattices, cette approche permet de conserver l’ordre de
précision de la LBM en vitesse. Dans le cas plus général d’une frontière courbe, son utilisation
entraine une perte de précision et une réduction de l’ordre de convergence [67, 73, 74].
28
Afin de réduire la perte de précision associée à l’imposition d’une vitesse non nulle sur une
frontière courbe, de nombreux travaux de recherche ont proposé des solutions basées sur des
approches différentes. Par mesure de simplicité, nous n’introduirons que les solutions facilement
applicables à une géométrie 3D complexe, c’est-à-dire les méthodes basées sur l’interpolation ou
l’extrapolation en fonction de la distance à la frontière [75-79] et les conditions aux limites
immergées [80-83].
2.3.2 Stratégies d’interpolation ou d’extrapolation
L’idée principale de ce type de stratégie de conditions aux limites est d’utiliser la distance à la
frontière pour augmenter la précision. Comme illustré dans la Figure 2.3, pour chaque duo de
lattices fluide-solide, la fraction
(appelée distance) est utilisée pour décrire exactement la
position de la frontière.
Figure 2.3: Illustration de la distance à la frontière (tirée de l'article 1)
La différence entre les méthodes se trouve dans la manière d’utiliser les distances
pour
améliorer la précision. Par exemple, la méthode d’extrapolation « extrapolation method » EM
[78] l’utilise pour calculer des populations d’équilibre et de non équilibre fictives au nœud solide
. Pour ce faire, la fonction d’équilibre est calculée avec une vitesse et une masse volumique
extrapolées à partir des nœuds voisins en
et
, la partie de non équilibre étant directement
extrapolée à partir des voisins (voir l’article 1 pour le détail des calculs). La définition des
populations de non équilibre est basée sur la décomposition présentée dans l’équation (2.13) et
réécrite de la manière suivante :
29
(2.44)
Une autre approche, proposée par Bouzidi, Firdaouss et Lallemand [77], interpole les populations
manquantes à l’aide des populations de post-collision voisines. Ainsi, les populations sont
calculées comme :
⃗
⃗
(2.45)
(2.46)
Ces méthodes permettent une augmentation significative de la précision et une conservation de
l’ordre de convergence en vitesse de la LBM [74, 78, 84]. Cependant, il y a deux désavantages
qui sont la perte de la localité des calculs (utilisation des lattices voisines) et la difficulté de
calculer les distances dans le cas d’une géométrie complexe, comme une pale ou un milieu
poreux.
2.3.3 Conditions aux limites immergées
Historiquement développées pour les méthodes d’éléments et de volumes finis, les conditions aux
limites immergées (IBM) [10, 40, 85] utilisent des points de contrôle pour contraindre la vitesse
du fluide environnant. Comme illustré dans la Figure 2.4, en immergeant un ensemble de points
décrivant la forme d’un obstacle, il est possible de contraindre le fluide à avoir sa vitesse, sans
définir de frontière au domaine fluide. Ce type d’approche est très largement utilisé dans le cas
d’obstacles mobiles puisqu’il permet de ne pas avoir à rediscrétiser le volume fluide en fonction
du déplacement d’un objet.
Plusieurs adaptations ont été proposées pour la LBM [80-83] et sont basées sur deux étapes : une
interpolation pour calculer la différence entre la vitesse aux points de contrôle et la celle de
l’obstacle; puis une distribution de la correction aux lattices voisines à l’aide d’une force
volumique. La différence entre les méthodes se trouve dans la manière de redistribuer la
correction. Par exemple la IB-LBM [83] propose de corriger la vitesse en deux étapes, d’abord
dans le calcul des vitesses puis dans l’ajout d’une force volumique. Cette approche est basée sur
30
la définition de la force volumique proposée par Guo et al. [65] et couple l’ensemble des
corrections à l’aide d’un système matriciel (voir l’article 1 pour le détail des calculs).
Figure 2.4: Illustration d'une condition aux limites immergée
La technique de forçage « forcing technique », proposée par Derksen et al. [86], est basée sur une
correction temporelle de la force volumique
définie comme :
⃗
avec
et
des paramètres numériques déterminés empiriquement et ⃗
point de contrôle de coordonnées
La fonction d’interpolation
la vitesse interpolée au
.
est utilisée pour le calcul des vitesses aux points de contrôle
⃗
avec ⃗
(2.47)
∑ (
)⃗
(2.48)
les vitesses dans les lattices environnantes.
La redistribution des corrections
d’interpolation
sur les lattices environnantes utilise aussi la fonction
:
∑ (
)
(2.49)
31
Ces méthodes permettent donc de ne pas rediscrétiser le domaine en fonction du déplacement
d’un objet et de ne pas avoir à initialiser de nouvelles lattices fluides.
2.4 Mouvement d’une partie mobile
Lorsque le problème simulé comporte des objets mobiles, il est nécessaire de modéliser l’action
de leurs déplacements sur le fluide. Dans le cas des conditions aux limites immergées, ce
traitement est largement simplifié puisqu’il se limite à déplacer les points de contrôle. À
l’inverse, pour les autres stratégies de conditions aux limites (règle du rebond et méthodes
d’interpolation ou d’extrapolation), il est nécessaire de refaire la discrétisation en fonction du
déplacement des objets. Comme la LBM est appliquée à des maillages cartésiens structurés,
l’étape de discrétisation se limite à déterminer l’ensemble des nœuds situés à l’intérieur des
parties solides. Il est important de souligner que cette étape ne comporte pas de difficulté
majeure, même dans le cas de géométries complexes. À l’inverse, l’initialisation des lattices
nouvellement fluides, illustrée dans la Figure 2.5, est une difficulté intrinsèque à la méthode. En
effet, de la même manière qu’il est difficile d’imposer une condition aux limites macroscopique,
les initialisations souffrent de la nature mésoscopique des populations.
Figure 2.5: Illustration de nouvelles lattices fluides après mouvement d'un objet
Afin de déterminer les populations des lattices nouvellement fluides, plusieurs façons sont
envisageables. Par exemple, il est possible d’interpoler les vitesses et les masses volumiques des
lattices voisines, puis de les injecter dans la fonction d’équilibre définie par l’équation (2.33). De
32
même, il est possible d’interpoler directement les populations voisines. Ces différentes approches
ont un niveau de précision équivalent, comme rapporté dans [87], mais présentent le désavantage
de générer des oscillations en vitesse et en masse volumique autours des objets mobiles [88].
Afin de réduire ces oscillations, il est préférable d’utiliser une initialisation avec un terme
correctif basé sur les populations de non équilibre « non-equilibrium corrected refill » [88].
Cette approche nécessite d’abord de choisir une direction d’extrapolation
, qui est ensuite
utilisée pour déterminer une vitesse ̃ et une masse volumique ̃ extrapolées. À partir de ces
données, les populations sont calculées comme :
̃
̃ ̃
(
)
(2.50)
Ainsi, l‘initialisation des nouvelles populations est corrigée avec les populations de non équilibre
des lattices voisines. Cette correction permet de réduire largement les oscillations et d’améliorer
le champ de vitesse autour de l’obstacle [88]. Il est important de souligner que l’étape de
sélection de la direction d’extrapolation
n’est pas générique. En effet, il n’existe pas de
formule pour calculer cette direction et sa détermination est laissée à la discrétion de l’utilisateur.
2.5 Intégration des modèles rhéologiques
Comme nous l’avons vu dans la section 1.2.2, l’expression du tenseur des contraintes, dans le cas
de modèles non newtoniens sans effet d’élasticité, prend la forme suivante :
̇
Dans ce cas, la viscosité
est fonction du taux de cisaillement ̇
(2.51)
√
ce qui implique que sa
valeur n’est plus constante dans le temps et dans l’espace. La dépendance de la viscosité en
fonction du taux de cisaillement est définie par le modèle rhéologique utilisé. Afin d’intégrer ce
modèle dans le schéma LBM, il suffit de modifier le calcul du facteur de relaxation. Son
expression devient alors :
̇
avec
calculé grâce l’équation du modèle.
̇
(2.52)
33
Il est important de noter que cette modification n’altère en rien le développement présenté dans la
section 2.1, c’est-à-dire que toutes les propriétés de la méthode LBM sont conservées. Afin de
conserver la localité complète de la phase de collision, le calcul du tenseur du taux de
déformation
n’est pas calculé à partir des vitesses des lattices voisines. En effet, il est possible
d’utiliser une expression différente du tenseur
∑(
Cependant, comme dans l’équation (2.53)
Ainsi, la détermination de
[62, 89] basée sur les populations :
)
(2.53)
dépend de , l’équation (2.52) devient implicite.
devrait être réalisée à l’aide d’une méthode de Newton-Raphson ou
de point fixe. Comme cette étape doit être appliquée dans chaque lattice et à chaque itération, le
coût en temps de calcul peut devenir prohibitif. Le calcul du facteur de relaxation
est donc
généralement traité avec une approche explicite, c’est-à-dire que la valeur utilisée dans l’équation
(2.53) est celle au pas de temps précédent [14, 90]. De cette manière le traitement de la rhéologie
se fait facilement avec la LBM.
Par rapport à une simulation LBM newtonienne, l’intégration d’un modèle rhéologique
complique la procédure de simulation. En effet, comme la variation totale de viscosité est
représentée par une variation du facteur de relaxation , celui-ci doit rester dans une plage qui
conserve la convergence du schéma. D’après nos tests, cette plage se trouve entre
la valeur de
et ,
étant dépendante du problème simulé. Par exemple, dans le cas d’un modèle
rhéofluidifiant (ou pseudo-plastique), il est nécessaire de choisir une valeur de
̇
, à taux
de cisaillement nul, assez grande pour que la valeur du facteur de relaxation à taux de
cisaillement maximal
̇
̇
n’entraine pas la divergence du schéma. En pratique, cette
condition limite la vitesse maximale simulable au travers du nombre de Mach, défini par
l’équation (2.38). En effet, comme
vitesse de lattice
(
|⃗ |
√
, une augmentation du facteur de relaxation réduit la
. Comme il est nécessaire de conserver un nombre de Mach faible
), cela peut réduire drastiquement la vitesse maximale | ⃗ | pour laquelle la solution
à un sens.
Plusieurs modèles rhéologiques ont ainsi été appliqués avec succès dans la LBM. Pour les
modèles sans viscoélasticité, on notera l’utilisation de modèles de loi de puissance [15, 16], de
34
Carreau-Yasuda [17, 18], de Bingham [27, 90] et de Casson [17, 90]. Il est important de
mentionner que dans le cas des modèles viscoélastiques, il est nécessaire d’ajouter une équation
supplémentaire afin de transporter les termes du tenseur de conformation [14, 91].
2.6 La LBM appliquée aux cuves agitées
Dans le cas de la méthode de Boltzmann sur réseau, ce sont principalement des turbines Rushton
qui ont été simulées en régime turbulent. La géométrie est modélisée à l’aide de la technique de
forçage [86] présentée à la section 2.3.3 et la turbulence est simulée directement (DNS) ou traitée
grâce à un modèle de simulation aux grandes échelles « Large Eddy Simulation (LES) » [86],
suivant les nombres de Reynolds atteints. Les simulations newtoniennes [86, 92, 93] ont permis
de valider l’utilisation de la LES dans ce contexte, par comparaison avec des données
expérimentales, des résultats de simulations numériques directes (DNS) et d’équations de
Reynolds moyennées (RANS). Dans le cas non newtonien, la simulation d’un liquide
thixotropique [54, 94] a permis d’étudier l’impact du nombre de Deborah sur la forme de
l’écoulement. On notera aussi la simulation du mélange de deux liquides avec des masses
volumiques et des viscosités différentes, à l’aide d’une méthode de type 1-fluide [95, 96].
Dans le cas des suspensions solide-liquide, l’utilisation d’une approche de type particules
discrètes (ou Euler-Lagrange) [26] a permis l’étude des temps de séjour [97] et de la dissolution
des particules [98]. L’impact de la rhéologie sur l’évolution de la suspension a aussi été étudié
pour des liquides de Bingham [27]. Cette approche de simulation Euler-Lagrange a aussi été
étendue à la simulation de suspensions gaz-liquide [99, 100]. Il est à noter qu’un autre groupe de
recherche a appliqué une méthode de Boltzmann sur réseau avec des particules colorées (c’est-àdire avec l’ajout d’une équation de transport d’une concentration) afin d’étudier le mélange d’un
système simplifié [101].
À partir de cette revue de littérature, il ressort que la LBM est une méthode efficace pour la
simulation du mélange turbulent. Autrement dit, lorsqu’elle est couplée à un modèle LES, elle
permet de reproduire globalement la création et le transport des tourbillons au sein d’un
mélangeur industriel. Dans le cas du mélange visqueux, ce sont les interactions entre la pale
raclante et la cuve ainsi que la forme de la pale qui génèrent le mélange dispersif et distributif.
Aussi, le traitement de la géométrie et l’imposition des conditions aux limites doivent être assez
précis pour prédire correctement les performances du système de mélange.
35
CHAPITRE 3
ORGANISATION DE LA THÈSE
La simulation des écoulements visqueux en cuves agitées requiert un traitement précis de la
géométrie, de la modélisation de l’impact des parties mobiles sur le fluide et de l’intégration des
modèles rhéologiques. Dans ce cas, l’utilisation des méthodes classiques de la CFD entraine
généralement un compromis entre la précision et les temps de calcul, limitant ainsi les
possibilités de simulation d’écoulements polyphasiques ou à rhéologies complexes. À l’inverse,
la méthode de Boltzmann sur réseau (LBM) permet un traitement simplifié des géométries
complexes et une intégration naturelle des modèles rhéologiques sans effet d’élasticité.
Cependant, la définition des conditions aux limites et le traitement de parties mobiles n’étant pas
triviaux, il est nécessaire de déterminer les erreurs qui peuvent être générées par ces deux aspects.
Finalement, comme l’intégration des modèles rhéologiques est simplifiée, il est intéressant
d’étudier les limites atteignables en termes de complexité.
Afin d’atteindre l’objectif principal, les objectifs spécifiques sont :

Comparer les différentes stratégies pour l’imposition des conditions aux limites en LBM;

Déterminer la précision de la LBM pour des écoulements visqueux dans des géométries
complexes;

Caractériser l’applicabilité de la LBM pour des rhéologies complexes sans effets
d’élasticité.
Afin de déterminer la stratégie la plus appropriée, l’article 1 présente une comparaison de trois
stratégies de conditions aux limites couramment utilisées dans la littérature et facilement
applicables à des géométries complexes 3D. D’abord, leurs impacts sur la précision de la LBM
sont quantifiés à l’aide de la simulation de l’écoulement de Couette cylindrique. Cette étape
permet aussi l’étude des paramètres numériques associés à chaque stratégie. La comparaison est
ensuite étendue à la simulation de deux systèmes de mélange utilisant une ancre et un Paravisc. À
l’aide d’un repère Lagrangien basé sur la pale, les trois stratégies de conditions aux limites sont
comparées sur la qualité de prédiction de la consommation de puissance et de la capacité de
36
pompage. Les impacts des trois stratégies sur l’efficacité de la LBM pour le calcul parallèle sont
aussi quantifiés.
L’article 2 s’intéresse à la précision de la méthode de Boltzmann sur réseau dans le cas de
géométries complexes. Afin de maximiser la précision, seule la stratégie de conditions aux
limites basée sur l’extrapolation est utilisée. Tout d’abord, la simulation de l’écoulement de
Couette cylindrique permet d’étudier la sensibilité de la méthode à des modifications mineures
sur les frontières du domaine de simulation. Ensuite, l’écoulement au sein du système de mélange
Maxblend est simulé pour différentes configurations. Les prédictions de consommations de
puissance et de capacités de pompage sont comparées avec les données expérimentales et
numériques disponibles dans la littérature. Premièrement, à l’aide d’un repère Lagrangien basé
sur la pale, l’impact du dégagement au fond de la cuve est caractérisé. Deuxièmement, la
simulation de la configuration avec contre-pales, nécessitant l’utilisation du repère naturel,
permet d’étudier la précision de la méthode dans le cas d’une pale en rotation (c’est-à-dire avec
une partie mobile).
Le dernier article traite de l’impact de la rhéologie sur les mécanismes associés à la création du
pompage dans le système de mélange Maxblend. D’abord, dans le cas newtonien, l’évolution de
la capacité de pompage en fonction du régime d’écoulement est comparée avec les données
expérimentales et numériques disponibles dans la littérature. Cette étape permet d’étudier la
transformation de l’écoulement secondaire (composantes radiales et axiales) dans le régime
laminaire profond. L’étude est ensuite étendue à la simulation d’un liquide fortement
rhéofluidifiant (indice de puissance de 0.05) à l’aide d’un modèle de Carreau-Yasuda. De cette
manière, l’impact de la rhéologie sur la transformation de l’écoulement secondaire est caractérisé.
Cette étape permet aussi de déterminer les limitations associées à l’utilisation de la LBM dans le
cas de rhéologies complexes (fortement rhéofluidifiantes dans ce cas).
37
Les contributions scientifiques associées à ce travail sont :

Une comparaison de stratégies pour l’imposition des conditions aux limites dans le cas de
géométries complexes;

Une étude de la précision de la LBM pour la simulation de la rotation d’un objet de forme
complexe;

La simulation de l’agitation d’un liquide fortement rhéofluidifiant (indice de puissance de
0.05).
Ce travail a nécessité de développement complet d’un code LBM tridimensionnel dont les
calculs peuvent être effectués sur un grand nombre de processeurs. Les particularités du code
sont relatives à sa capacité à définir des conditions aux limites sur des géométries complexes
ainsi qu’à prendre en compte la rotation des objects. Il est à noter que les résultats présentés
dans l’article 1 ont été obtenus grâce à un code LBM développé par D. Vidal.
38
CHAPITRE 4
COMPARAISON DES DIFFÉRENTES STRATÉGIES
POUR L’IMPOSITION DES CONDITIONS AUX LIMITES EN LBM
4.1 Présentation du premier article
Accepté dans Computers and Fluids le 26 Décembre 2012.
Auteurs : Vincent Stobiac, Philippe Tanguy, François Bertrand.
Dans cet article, trois stratégies d’imposition de conditions aux limites sont comparées dans le
cadre de la simulation d’écoulements visqueux en cuves agitées. Les stratégies ont été
sélectionnées parmi les approches les plus couramment utilisées dans la littérature : la méthode
du rebond modifié (MBB), la méthode d’extrapolation (EM) et la méthode de conditions aux
limites immergées (IB-LBM), chacune de ces stratégies étant représentative d’une classe de
conditions aux limites possédant des propriétés similaires. Tout d’abord, l’écoulement de Couette
cylindrique a été simulé afin de vérifier l’impact des trois stratégies sur l’ordre de convergence du
schéma LBM. Il ressort que seule la méthode d’extrapolation est capable de conserver la
précision du deuxième ordre pour la convergence spatiale de la vitesse. Ensuite, à l’aide d’un
code LBM massivement parallèle, les simulations sont étendues au cas des écoulements dans des
mélangeurs raclants (ancre et Paravisc). Le traitement de la géométrie est simplifié par
l’utilisation d’un repère Lagrangien fixe par rapport à la pale. Ainsi, les trois stratégies sont
comparées sur la convergence spatiale des prédictions de consommation de puissance et de
capacité de pompage. Si tous les résultats LBM sont globalement en accord avec les données
expérimentales et numériques disponibles dans la littérature, il ressort que la condition aux
limites immergées semble moins adaptée aux mélangeurs raclants. À l’inverse, la méthode du
rebond modifié, malgré sa simplicité, permet d’obtenir des prédictions équivalentes à celles de la
méthode d’extrapolation. Finalement, l’impact des trois stratégies sur les performances en calcul
parallèle est aussi quantifié. Cela montre que seule la stratégie basée sur une condition aux
limites immergées perturbe les propriétés de la méthode de Boltzmann sur réseau.
Note : Dans cet article, le terme « Lagrangian frame of reference » réfère à un repère tournant
fixé sur la pale.
39
4.2 Boundary conditions for the lattice Boltzmann method in the case of
viscous mixing flows
4.2.1 Summary
This paper examines the performance of different boundary condition strategies for the lattice
Boltzmann simulation of industrial viscous mixing flows. Three different strategies were chosen
from the most popular approaches, which are the staircase approaches (bounce-back rules), the
extrapolation or interpolation methods, and the immersed boundary methods. First, the order of
convergence of the selected methods to impose boundary conditions is verified on the 3D Couette
flow. This work clearly shows that only the extrapolation method is capable of preserving the
second order accuracy of the lattice Boltzmann method. Second, a highly parallel LBM scheme is
used to simulate fluid flow with close-clearance mixing systems in the Lagrangian frame of
reference. The convergence rates obtained with the different boundary condition strategies are
compared on the basis of two characteristic mixing numbers, the power number and the pumping
rate. The results agree well with experimental data and finite element simulation results, and
surprisingly enough, the modified bounce back rule provides a reliable accuracy despite its
simplicity. Finally, the impact of the boundary condition strategies on the workload balance and
the memory usage is analyzed. It appears that only the immersed boundary condition strategy
modifies the parallel efficiency of the lattice Boltzmann method, yet no significant effect is
observed on the memory usage.
4.2.2 Introduction
Mixing of fluids in stirred tanks is a common activity in many industries such as chemical,
pharmaceutical and food processing. During the batch process, the mixing is driven by impellers
which are chosen relative to the process characteristics and the fluid properties. Given that there
are a large number of important parameters to control the mixing efficiency (bottom clearance,
number and position of impellers, speed ratio, etc), its optimisation is a difficult task not to
mention costly on an experimental level. In order to reduce this cost, computational fluid
dynamics (CFD) has been largely used to compare the performance of mixing systems [41, 43,
45] and to optimize mixing configuration [36, 42, 46, 52, 102, 103]. Another distinguished
advantage of CFD is its capacity to access physical variables (velocity, density and shear stress,
40
etc) inside a system without disturbing the flow. In addition, the simulated flow field can be used
to compute characteristic macroscopic values such as mixing time [36, 41, 46, 103], power
consumption [12, 13, 36, 41, 42, 44, 46, 48, 51, 52, 103-105] and pumping capacity [13, 41, 46,
106].
One key aspect of the simulation process is the modeling of the action of impellers on the flow,
which requires the imposition of a velocity profile along moving surfaces. This task is
complicated due to the unfeasibility of remeshing after each time step. Thus, what is required for
a fixed discretization is a technique capable of imposing different velocity profiles at different
places. Several solutions with entirely different approaches have been proposed in the literature.
The sliding mesh (SM) technique widely used with baffled vessels decomposes the domain into
two coaxial subdomains, one in the impeller region and its complementary which extends to the
tank wall and contains the baffles. This technique is limited to impeller motions based on one
single axis and requires communication between the different parts. In order to describe more
complex motions, i.e. rotations around several axes, the classes of fictitious domain methods [10]
and immersed boundary methods [9] consider a fully fluid domain in which the fluid velocity is
constrained at various control points representing the immersed solid parts. Finally, for fully
periodic flow, the steady state solution can be computed by employing the Lagrangian frame of
reference [37]. In this case, impellers are considered to be fixed and then simply defined by a
homogeneous Dirichlet (i.e. zero velocity) boundary condition. This approach is limited to
configurations without baffles and with impellers positioned along the same axis, with the same
velocity. For the particular case of viscous mixing, no baffles are generally required so that this
approach can generally be used, which greatly simplifies the simulation process. For this reason,
this strategy was chosen in this work.
Although different CFD methods have been applied to simulate hydrodynamics in stirred tanks,
the finite element method (FEM) appears to be the most popular approach for the laminar and
transition regimes, as a lot of mixing configurations have been simulated. Using the Lagrangian
frame of reference, the hydrodynamics of the Paravisc impeller [41], double helical ribbons
(DHR) [12, 37, 43], the Maxblend impeller (without baffles) [13, 36, 43, 44], dough mixers [42]
and the Rushton turbine (without baffles) [107] have been analysed. These simulations have
allowed to quantify the mixing of Newtonian [12, 13, 36, 37, 41-44, 107] and non-Newtonian
flows [12, 13, 41, 42], the power consumption [12, 13, 36, 41, 42, 44], the pumping capacity [13,
41
41] and the mixing time [36, 41] in specific mixing systems. For planetary mixers, the use of the
Lagrangian frame of reference is no longer possible and more complex boundary condition
strategies are required, as explained in the previous paragraph. For instance, a three-blade
planetary mixer has been simulated [44] using the so-called virtual finite element method
(VFEM) [40], which belongs to the category of fictitious domain methods. In the same manner, a
two blade planetary mixer has been simulated [45] by a simplified version of the VFEM,
implemented in FIDAP (ANSYS). For the simulation of a planetary pin mixer [108], another type
of fictitious domain method, called the mesh superposition technique (MST) [109], has been
applied. In order to simulate baffled mixing configurations or to access unsteady behaviour, the
same techniques have been applied. For instance, the VFEM has been used for the Maxblend
impeller (baffled and unbaffled) [13, 36, 44] as well as for coaxial mixers comprised of an anchor
and a Rushton turbine [103] or a system of rods [105]. For the Superblend [46], comprised of a
Maxblend and a DHR impeller, the SM technique was employed for the Maxblend, coupled with
the VFEM for the DHR.
As for mixing flows in turbulent regime, a lot of simulations have been performed with the finite
volume method (FVM) in the case of tanks with Rushton turbines [48-51]. The action of the
impeller on the fluid is generally achieved by the SM technique or by similar approaches such as
the multiple reference frames (MRF) method, which can be viewed as a simplified version of the
SM technique only valid for steady-state flows. These simulations have permitted to analyze the
turbulent kinetic energy [50, 51], the energy dissipation rate [51] and the power consumption [48,
51] for various mixing configurations. Other types of impellers have also been investigated. For
instance, the power consumption of a pitched turbine has been analyzed [104] using a hybrid
method based on the so-called blade element theory. It must be pointed out that the FVM, along
with the SM technique, has also been applied in the laminar and transition regimes. We can cite,
for instance, an investigation of the hydrodynamics in a coaxial mixer comprised of an anchor
and a pitched blade turbine [52].
According to this literature review, it appears that the FEM and the FVM have been successfully
used in a large number of mixing configurations and have been generally capable of accurately
describing the hydrodynamics of a single phase in these mixing systems, particularly in local
regions such as the area near the impeller. However, many industrial processes involve more than
one phase, a typical example being the mixing of solid particles in a non-Newtonian polymeric
42
solution. The simulation process is then complicated by the coupling term between the phases
and the rheological model, which generally leads to an increase in the total number of unknowns
and a complexification of the system of equations to be solved. In addition, these methods often
require the use of complex algorithms for parallel computations (see [46] for an example of FEM
on parallel computers) and nevertheless necessitate large computational times for simulating real
industrial mixing flows, resulting in a compromise between cost and accuracy. In order to reduce
the complexity of the simulation procedure, it is more and more admitted that other modeling
paradigms are needed.
The lattice Boltzmann method (LBM) is a powerful numerical method which has been
successfully used in many research areas, such as porous media [1, 110], solid-fluid suspensions
[53, 72, 111] and multiphase systems [53, 111, 112]. This method is based on the kinetic theory
of gas and has been the topic of several researches in the last decade. Contrary to classical CFD
approaches such as the finite element and volume methods, the Navier-Stokes equations are not
directly solved with the LBM. In this case, the solution of the Navier-Stokes equations is
recovered by applying a simple recurrence equation made up of two steps: the collision and the
propagation steps of particles along the bonds of a lattice grid representing the computational
domain. As a result, the only required computations are done during the collision step and are
simple and fully local. For the propagation step, the information is directly translated from one
lattice node to the neighboring lattice nodes. Thus, when this method is applied on a Cartesian
structured mesh, the parallelization of the scheme is straightforward and the parallel efficiency is
highly favorable.
The main difference between the LBM and classical CFD methods is the physical scale. By
considering the flow at the mesoscopic scale, the primary variable of the LBM is a population of
fluid particles. This change of scale presents several advantages, such as an easier way to define
turbulence models [86, 113, 114], to model non-Newtonian flows [54, 115, 116] or to simulate
multiphase flows [53, 111, 112], yet unfortunately it reveals a difficulty when imposing boundary
conditions on macroscopic variables. In the case of homogeneous Dirichlet boundary conditions
(i.e. zero velocity) on a straight wall, this limitation can be circumvented by using the so called
bounce-back rule [58, 71, 72], although from a more general perspective, the imposition of nonzero velocities along arbitrarily curved surfaces is still an open topic (this point will be discussed
in section 4.2.4). Nevertheless, using the adaptive force field technique [80], which belongs to the
43
class of the immersed boundary methods, the LBM has been successfully used to simulate the
hydrodynamics of a Rushton turbine in the turbulent regime [86]. This simulation process has
been used to simulate Newtonian [86, 117], Bingham [27] and thixotropic [54] fluids in order to
quantify the turbulent kinetic energy [86, 117], the apparent viscosity [27] and the volume of
caverns [27]. From these results, the LBM appears to be a promising method for the simulation of
mixing systems. However, in the case of viscous mixing, close clearance impellers are generally
used, and this smaller wall clearance leads to stronger impeller-wall interactions, which can only
be captured by a geometrically flexible and numerically accurate treatment of the boundary
conditions.
In order to select the most suitable boundary condition strategy, several comparisons have
already been made for simple geometries. For example, the capacity to generate a zero slip
velocity and to conserve the second order accuracy in space of the LBM have been studied in the
case of Poiseuille and plane Couette flows [73]. The accuracy of the velocity field and the
conservation of the total mass have also been investigated for an impulsively started circular flow
[118]. In addition, the accuracy of the imposition of a no-slip boundary condition and the
influence of the inherent parameters have been assessed for the 2D circular Couette flow [74].
From these investigations, it appears that the choice of the boundary condition strategy can
highly impact the velocity at the boundary, the global accuracy of the velocity field and the order
of convergence of the LBM. The objective of this work is then to assess the suitability of
different strategies for imposing boundary conditions for the lattice Boltzmann method (LBM) in
the case of the 3D simulation of viscous mixing flows. To our knowledge, no such comparison
has been made for such complex 3D flow problems. The conclusions of this work are not limited
to mixing flows and may be applicable to other types of geometrically complex flow problems.
The organization of the paper is as follows: in Section 4.2.3, the LBM is recalled and the
specifications of the implementation used in this work are described. In Section 4.2.4, three
commonly used strategies for imposing boundary conditions are explained. In addition, the
modifications required for their implementation into a parallel LBM scheme are presented. In
Section 4.2.5, the three-dimensional Couette flow is used to analyze the order of convergence of
these boundary condition strategies. The impacts of the parameters inherent to these techniques
are also investigated. In Section 4.2.6, a highly parallel lattice Boltzmann scheme is used to
simulate the hydrodynamics of the anchor and the Paravisc (Ekato GmbH) impellers. The
44
accuracy of the different strategies is quantified using two macroscopic characteristic numbers,
which are the power number and the pumping rate. Finally, in Section 4.2.7, the parallel
performance of every boundary condition strategy is discussed and its impacts on the workload
balance as well as the memory usage are analyzed.
4.2.3 The lattice Boltzmann method
The lattice Boltzmann method [58, 61, 119, 120] is based on the kinetic theory of gas and comes
from the discretization in space, velocity and time of the Boltzmann equation. Only a brief
presentation of the method is given here and the reader is referred to [58] for more detailed
explanations. Basically, the idea is to obtain the macroscopic flow from the behaviour of a large
number of microscopic particles that propagate and collide along the bonds of a lattice grid
representing the computational domain. Thus, the primitive variable
is mesoscopic and
describes a population (or a probability density function) of particle at position , with velocity
, at time t. The macroscopic variables can be easily recovered from the populations as the
following relations define the density and the velocity:
(4.1)
∑
(4.2)
⃗
where
∑
is the total number of populations per lattice.
With this method, the solution of the Navier-Stokes equations can be recovered by successively
applying the following collision and propagation steps:
(
)
(4.3)
(4.4)
where
is the collision operator, as discussed below.
45
This means that, after collision, the populations are propagated along the direction defined by
during a time step
. Here,
corresponds to the time just after the collision step. The lattice
grid discretization, which is Cartesian and regular, is based on the lattice spacing
. In the
present work, the D3Q15 lattice [63] is chosen and the velocities are defined as follows:
{
(4.5)
From these definitions, it can be observed that the collision step only requires local computations,
and the propagation step is simply a translation of information along a structured grid. This is
why the LBM can be parallelized efficiently.
4.2.3.1 The Bhatnagar-Gross-Krook collision model
The BGK model [60, 63] is the most commonly used collision model. It is based on a single
relaxation towards the equilibrium population
, using a relaxation parameter . The
collision operator in (4.3) is defined as follows:
(
)
[
(4.6)
]
and the equilibrium population is given by:
⃗
[
where
by:
⃗ ⃗
]
(4.7)
is the speed of sound defined as:
√
and
⃗
(4.8)
are the weights of the corresponding populations. For the D3Q15 lattice, they are given
46
(4.9)
{
The relaxation parameter is related to the fluid viscosity
by the following relation:
(4.10)
With diffusive time scaling (i.e.
), this model has proven to be second-order accurate in
space. The mathematical demonstration is derived from an asymptotic analysis [67] or a
Chapman-Enskog expansion [121] and, under the hypotheses of small Mach and Knudsen
numbers, the solution of the Navier-Stokes equations is recovered. In the case of the LBM, these
dimensionless numbers are defined as follows [58]:
‖⃗ ‖
(4.11)
(4.12)
where
is a characteristic length.
It must be pointed out that although more advanced models are available, the MRT [122] and the
regularized LBM [123] schemes for instance, the BGK model was retained for the present study
since it provides a sufficient level of accuracy, while offering a cost reduction factor of 10 or 20
percent [124] over these other models.
In this work, the relaxation parameter was set to one, so that Eq. (4.3) and (4.4) reduce to:
(4.13)
The use of this value presents the advantage to reduce the memory cost, as the populations are
equal to the equilibrium populations. In this case, the populations do not need to be stored, as the
equilibrium function only requires the velocity vector and the density. This means that only four
47
real numbers per lattice need to be stored instead of fifteen. By reducing the amount of data
required to apply the LBM scheme, the refinement possibilities are enhanced [2] so that the
emphasis can be put on the lattice spacing rather than on the relaxation parameter. In fact, by
fixing this parameter, the accuracy is not always optimal (an investigation of the influence of the
relaxation parameter on accuracy can be found in [74]), although the computational time and the
memory usage greatly decrease, hence a higher number of lattices can be used.
4.2.3.2 Parallel computation strategy
As mentioned in the introduction, in the special case of viscous mixing, the distance between the
tank and the impeller is actually quite small compared to the size of the tank. As an accurate
description of the flow inside this small wall clearance is crucial for the understanding of the
mixing process, the discretization in this zone must be refined enough to capture the flow
correctly. Thereupon, when applied to the rest of the volume, the discretization of the fluid zone
leads to a substantially high number of lattices. Coupled with the fact that the LBM is based on
an explicit scheme, which requires the use of small times steps, the parallelization of the
computation is mandatory for performance.
The parallelization of the LBM, which relies on domain decomposition techniques, is briefly
examined here and the reader is referred to [2] for more details. To minimize the memory usage,
the solid lattice nodes (i.e. the lattices inside a solid part) are removed from the computational
domain, leaving only the fluid lattice nodes to be distributed amongst the processors in such a
way that the workload balance is optimal. This also minimizes the amount of data to be
communicated before every new time step. Unfortunately, this further complicates the interprocess communications since the domain is no longer continuous, which implies that the
advantages of a regular Cartesian mesh are often lost (i.e. straightforward determination of the
neighbours of a given lattice). Moreover, as mentioned in the previous paragraph, by setting the
relaxation parameter to one, the population does not need to be stored for the next iteration step,
as the velocity and the density are sufficient to compute the equilibrium function. This means that
the amount of data that must be transferred between processors is reduced. By virtue of these
reductions of memory usage and efficient domain decomposition strategies, our code has been
successfully utilized for the direct numerical simulation of porous media flows [1] using as many
as 3.4 billion lattice nodes, spread among 256 cores (3.6 GHz Intel Xeon) for a computational
48
time lower than 12 hours. Such high parallel performance is possible thanks to the intrinsic
locality and simplicity of the collision-propagation procedure. Consequently, one should keep in
mind that boundary condition strategies can strongly affect the workload and increase the
memory usage. We will come back to this in Section 4.2.7.
4.2.4 Boundary condition strategies
With the LBM, the macroscopic variables (velocity and density) are easy to obtain from the
mesoscopic variable (population). However the inverse transformation is generally not possible.
One major drawback of this difficulty concerns the definition of boundary conditions, as
simulations generally require the imposition of velocity profiles along boundaries of the domain.
From a practical point of view, this difficulty can be explained by the greater number of
populations than the number of macroscopic variables per lattice, so that a Dirichlet boundary
condition on velocity cannot be directly applied. As a result, the boundary conditions modify the
populations of the corresponding lattices in order to tend to the required velocity.
In the case of homogeneous (i.e. zero) Dirichlet boundary conditions, the so-called half-way
bounce-back rule [58, 71, 72] can be easily applied to the unknown populations. This rule is
given by:
(4.14)
where
corresponds to the direction opposite to
be positioned at
and the boundary (e.g. wall) is considered to
(see Fig. 4.1).
This rule simply reflects the outgoing populations obtained during the propagation step in order
to define the unknown populations for the next collision step. Thanks to its simplicity, this
boundary condition strategy has become the most popular approach for imposing no-slip (zero)
velocities, as it allows for the handling of fluid domains with complex shapes without losing the
locality of the scheme. Unfortunately, this approach is second-order accurate only in the case of
straight walls perpendicular to one axis of the simulation reference frame and positioned exactly
in the middle of two consecutive layers of lattices. This limitation generally implies a reduction
to the first order of accuracy in space in the case of curved walls [73, 74].
49
Figure 4.1: Populations used for the half-way bounce-back rule
The imposition of non-zero velocities along a complex surface is still an open topic, and several
authors have proposed various boundary condition strategies based on different approaches. In
order to classify them, the different methods can be split into three categories: The staircase
approaches [72, 125], the interpolation or extrapolation methods [75-79, 118, 126], and the
immersed boundary based methods [80-83].
In this study, three boundary condition strategies were considered: the modified bounce-back
(MBB) rule proposed by Ladd [72], the extrapolation method (EM) proposed by Guo et al. [78]
and the enforced immersed boundary-lattice Boltzmann method (IB-LBM) proposed by Wu et al.
[83]. These methods are each representative of one category, in the sense that the MBB,
independently of the shape of the boundary, uses the outgoing populations to define the unknown
ones, the EM uses the distance between fluid nodes and solid parts to improve the accuracy of the
boundary conditions, and the IB-LBM uses a group of control points and a corresponding body
force to constrain the velocity of neighboring fluid nodes in order to mimic the motion of solid
parts.
There is mention of other approaches in the literature, but they appear to be too complicated to
tune and use in the case of mixing systems or other processes of similar or greater complexity
(e.g. volumetric method [127, 128]) or they require a modification of the relaxation parameter
(e.g. adaptive relaxation time boundary conditions [129]).
4.2.4.1 The modified bounce-back rule
The modified bounce-back (MBB) rule [72] is a classical half-way bounce-back method in which
a second term, proportional to the solid part velocity ⃗ , is used to modify the population density:
50
⃗
where
corresponds to the direction opposite to
positioned at
(4.15)
and the solid part is considered to be
(Fig. 4.2).
Figure 4.2: Populations used for MBB
The discretization of the solid part may be inexact, as the boundary is considered to be located
midway between every couple of fluid-solid nodes. In particular, the discretization of a curved
wall generates a staircase profile (see Fig. 4.3). This inaccuracy generally leads to the loss of the
second order accuracy in space of the method [73, 74]. Moreover, the imposed velocity ⃗
may be treated inaccurately, as only its projection onto the direction
itself
is taken into account,
which means that this boundary condition strategy approximates the action of the solid part on
the fluid.
However, this method is quite easy to apply for complex three-dimensional domains. First, the
only requirement is to determine which lattice nodes are solid, and this task is facilitated by the
intrinsic regularity of the mesh. Second, as with the regular bounce-back rule, the locality of the
collision step is conserved so that there is no impact on the parallelization of the LBM itself.
Finally, the utilization of this approach adds no numerical variables, as every variable in the last
term of Eq. (4.15) is already known.
The localization of the solid nodes is easily achieved by determining which lattice nodes are
inside the solid components of the domain. In this work, a volumetric mesh is used to discretize
these solid parts into elements (or cells) and facilitate their localization. When these parts are
fixed, as in the case in this work, this procedure only needs to be applied once. This
preprocessing step is then negligible compared to the iterative process of the LBM.
51
Figure 4.3: Discretization with MBB
4.2.4.2 The extrapolation method
In the extrapolation method (EM) [78], the unknown populations and their velocities are
computed at solid nodes, the black crosses in Fig. 4.4, by an extrapolation procedure, using the
velocity and the populations of neighboring fluid nodes and the fraction of the intersected link
from the solid part.
Figure 4.4: Discretization with EM
52
This approach allows for a more accurate shape of the solid part (see Fig. 4.4), as the boundary is
no longer considered to be located midway between fluid and solid nodes (see Fig. 4.5). The
fraction
where
is computed as follows:
,
and
|
|
|
|
(4.16)
correspond respectively to the position of the solid node, the fluid node and
the intersection between the actual boundary and the line joining
and
(see Fig. 4.5).
Figure 4.5: Fraction to the boundary with EM
Let us first introduce the decomposition of a population into equilibrium and non-equilibrium
parts:
(4.17)
Using Eqs. (4.3) and (4.6), the population at solid node
required for the next propagation step
is then given by:
̃
As node
(
)
(4.18)
is solid, the two terms on the right-hand side are unphysical and must be obtained by
extrapolation. First, for the equilibrium part of Eq. (4.18), the regular equilibrium function is used
(see Eq. (4.7)):
53
̃
⃗
[
⃗
⃗
⃗
]
(4.19)
where the macroscopic variables are determined using the neighboring nodes. More precisely, the
density
is approximated by (
) and the velocity ⃗
is determined by an extrapolation
procedure. Following the formulation of the extrapolation parameters proposed by Suh et al. [74],
the velocity ⃗
is computed as follows:
⃗
where ⃗ , ⃗ and ⃗
⃗
⃗
⃗
correspond to the velocity at nodes
the extrapolation parameters
(4.20)
,
and
, respectively, and where
are given by:
{
(4.21)
(4.22)
{
(4.23)
{
Second, the non-equilibrium part of Eq. (4.18) is directly extrapolated from the neighboring
nodes:
(
where the extrapolation parameter
)
(
)
(4.24)
is given by:
{
(4.25)
54
Based on these definitions, it can be observed that this boundary condition strategy relies on four
different extrapolation parameters:
,
,
and . These parameters are local (i.e. different for
every lattice and every direction) and only depend on the fraction
(see Eq. (4.21), (4.22), (4.23)
and (4.25)). Henceforth in this paper, this formulation will be referred as the original EM
formulation.
As proposed by Suh et al. [74], the extrapolation parameters
,
,
and
can be considered
as free numerical parameters. They showed that, for a given geometry, one can optimize them in
order to minimize the error and/or improve the stability. By means of an analysis of the slip
velocity on a straight wall, they found the following relations:
(4.26)
(4.27)
or, using Eq. (4.26):
(4.28)
As a result, in such a case, only
and
can considered free. In that they are also global (i.e.
constant for all lattices and directions), one can choose values of these parameters that will
maximize the accuracy and/or the stability of the LBM scheme.
As previously mentioned, the relaxation parameter τ is fixed to one in this work. This implies that
the last term of Eq. (4.18) vanishes, and only the equilibrium part of the populations is needed.
Consequently, the extrapolation parameter
of the non-equilibrium part is no longer useful,
which means that the only free parameter is
. Throughout the rest of this work, this formulation
will be referred to as the free EM formulation.
It is worth noting that, for both the original and free EM formulations, extrapolation parts of the
boundary condition strategy, that is the computation of ⃗
of
using Eq. (4.24), imply that the boundary condition strategy is no longer local.
Nevertheless, by computing ⃗
first fluid node
using Eq. (4.20) and the computation
and
when treating the collision-propagation step of the
, no modification is required in the data structure of the LBM code.
55
Additionally, it eliminates the need to increase the amount of data communicated between related
processors, as the information transmitted for the next collision-propagation iteration is sufficient
to compute the boundary conditions.
In the same way as the MBB rule, the solid nodes need to be identified, therefore volumetric
meshes are also used to determine them by selecting all the lattice nodes situated inside one
element of the mesh. This boundary condition strategy also requires determining the fraction
for every couple of solid-fluid lattice nodes. Even if these fractions are not additional numerical
parameters, in the sense they are only related to geometrical aspects, their computation can be
rather time consuming. For simple solid parts, such as planes or cylinders, the computation can
be done analytically as the intersection between the line defined by
and the equation of such a
part leads to solving a linear or quadratic polynomial equation. In the case of geometrically
complex objects, such as impellers, analytical parameterizations of the boundary shape are rarely
available. In this work, the impeller surfaces are defined by surface meshes of boundary
elements, implying the use of a search algorithm to find the elements intersecting the related
directions defined by the couples of solid-fluid lattice nodes. In each case, the coordinates
of
the boundary, which are used in the fraction computation (Eq. (4.16)), are the intersection of the
line defined by
and the corresponding surface element of the mesh. Once again, in the case of
a Lagrangian frame of reference, the computation of the fraction only has to be done once since
the solid parts are fixed. The impact of the fraction computations on the computational time will
be addressed in Section 4.2.7.
Finally, we have observed that, for some specific configurations, the boundary condition can be
ill-defined with this technique. This occurs when two opposite directions for the same lattice
point towards solid walls. In this case, the extrapolated velocities (Eq. (4.20)) at the later solid
nodes cannot be computed as the velocity of the second fluid node, located at
, is not defined.
When this happens, the velocity of the solid part is directly used instead, without extrapolation.
4.2.4.3 The immersed boundary – lattice Boltzmann method
With classical CFD methods, immersed boundary methods consider a fully fluid domain in which
the fluid velocity is constrained at various control points representing the immersed solid part.
The boundary condition is then enforced by modifying the velocity of neighboring nodes in order
to tend to the required velocity. When using the LBM, the velocity is not a primitive variable; the
56
immersed boundary-lattice Boltzmann method (IB-LBM) [83] enforces a velocity ⃗
on control
points by means of a body force (see Fig. 4.6).
Figure 4.6: Example of control point distribution with IB-LBM
Let us first introduce the body force formulation proposed by Guo et al. [65], which can be
written as:
(
(4.29)
)
with
(
)
(
⃗
⃗
)
(4.30)
(4.31)
⃗
where the body force
∑
is now derived.
Following Eq. (4.31), the velocity can be decomposed as:
⃗
⃗
⃗
(4.32)
57
where the intermediate velocity ⃗ is given by:
⃗
so that the body force
(4.33)
∑
can be expressed as a function of the velocity correction ⃗ :
⃗
(4.34)
The body force is then proportional to the velocity correction ⃗ , which is a priori unknown but
can be related to the velocity correction ⃗ at the neighboring control points.
Let us now consider a group of
(
control points and their respective coordinates
)
where
coordinate. We associate with every control point
⃗ (
is the Lagrangian parametric
an unknown velocity correction
).
⃗ and
In order to couple the velocity corrections
⃗ , the Dirac function
is introduced,
which is approximated by [9, 85]:
{
(
(
This function can be extended to vector
(
)
(
| |
))
(4.35)
:
) (
) (
)
(4.36)
The velocity correction ⃗ at a lattice node can then be approximated as:
⃗
∑ ⃗ (
) (
)
(4.37)
58
where
is the arc length (in 2D) or the area (in 3D) of the boundary element inside of which
the control point is located (see Fig. 4.6 for an example in 2D).
The next step consists of relating ⃗ to the velocity at the control points ⃗ , the velocity at the
fluid nodes ⃗ , and the velocity of the solid part ⃗ . To do so, ⃗
at the control point
is
interpolated using the Dirac function of Eq. (4.36):
⃗ (
where
)
∑⃗
(
)
(4.38)
is the computational domain.
By substituting Eq. (4.32) into Eq. (4.38), the following relation is obtained:
⃗ (
)
∑⃗
(
)
(4.39)
∑ ⃗
(
)
or, using Eq. (4.37):
⃗ (
)
∑⃗
(
)
(4.40)
∑∑ ⃗ (
(
) ) (
(
))
(
)
As the goal of the method is to enforce the velocity of the solid part on the control points, ⃗ is
then replaced by ⃗ :
⃗ (
)
∑⃗
(
)
(4.41)
∑∑ ⃗ (
(
) ) (
(
))
(
)
59
Considering that Eq. (4.41) is valid for every control point, two vectors are next introduced, the
unknown correction vector ⃗ and the objective vector ⃗ :
⃗
⃗
⃗ (
⃗ (
)
)
∑⃗
[
]
(
(4.42)
[
)
]
(4.43)
The substitution of Eq. (4.42) and (4.43) into Eq. (4.41) leads to the following system of linear
equations:
⃗
⃗
(4.44)
where
∑
(
By definition the matrix
(
))
(
)
solely depends on the lattice spacing
[
]
(4.45)
and the coordinates of the
control points, which implies that this matrix only have to be computed once for a fixed solid
part. In addition, when applied to a large number of control points, the size of the system
becomes crucial. The conjugate gradient method is used in this work since it outperforms direct
methods both in terms of memory and CPU time when the size of the system becomes large,
which is often the case for industrial problems. Note that this part of the computation has not
been parallelized.
At this point, the boundary condition is entirely defined. Indeed, solving the system of equations
(44) yields the velocity correction
⃗
at the control points. The velocity correction
⃗ at the
lattice nodes is then computed through Eq. (4.37) and used to correct the velocity in Eq. (4.32).
Finally, the body force , which can be computed easily from to the velocity correction ⃗ with
Eq. (4.34), is applied in the collision step of Eq. (4.29). This procedure is summarized in the
following algorithm [83]:
60
1) Compute matrix with Eq. (4.45) (only at
in case of fixed objects);
2) Apply the collision-propagation scheme involving the body force , as defined in Eq.
(4.29) (
), and compute macroscopic variables and ⃗ with Eq. (4.1) and Eq.
(4.33);
3) Compute vector ⃗ defined by Eq. (4.43);
4) Solve the linear system of Eq. (4.44) to obtain the correction ⃗ at the control points and
use Eq. (4.37) to compute the correction ⃗ at the fluid nodes and Eq. (4.34) to compute
the body force ;
5) Correct the velocity at the lattices with Eq. (4.32) and compute the external force using
Eq. (4.30);
6) Compute the equilibrium function defined by Eq. (4.7);
7) Move the control points (in case of moving objects);
8) Repeat step 1 to step 7 until convergence is reached.
Owing to this approach, a velocity profile can be imposed at the surface of a solid part with the
use of a surface mesh. Unfortunately, because this method is intrinsically global, the locality of
the LBM scheme is not preserved. First, this method requires solving a system of equations,
which increases the computational time, and implies additional communications between
processors. The impact on the parallel efficiency and the memory usage will be addressed in
Section 4.2.8. Second, during the interpolation and spreading steps (Eqs. (4.38) and (4.37)), the
exchange of information between lattice nodes and control points is obviously not local. In order
to reduce the impact of these two steps, another approximation of the Dirac function is used in
this work [130]:
(
(
| |
√
| |
√
)
| |
)
| |
(4.46)
{
which reduces the support of the function from 4 to 3. The major advantage of this reduction is
that, for a given control point, the neighboring lattice nodes are the same as those required for the
propagation step, thus no modification in the data structure of the LBM program is required.
Finally, in order to prevent dimensional problem with the term
following definition:
of Eq. (4.37), we propose the
61
(4.47)
with
(4.48)
where
denotes the surface of the solid part and
In this definition, the term
is the total number of control points.
has the advantage to be related to the total number of control points
and the lattice spacing. This point will be highlighted in Section 4.2.5.2
4.2.5 Convergence study
In order to verify the order of convergence of the boundary condition strategies proposed in the
previous section, the three-dimensional Couette flow is analyzed. Given that this flow can be
seen as a simplified mixer (i.e. with a centered cylindrical impeller), the conclusions are of
interest with regard to mixing simulations. Moreover, its simplicity permits the analysis of the
boundary condition parameters and helps to select them properly for real industrial mixing
systems.
4.2.5.1 Convergence study
This problem has the advantage of possessing an analytical solution even if the boundaries are
curved. In this work, the inner cylinder is rotating and the outer is fixed (Fig. 4.7).
Figure 4.7: Couette flow geometry
62
In the case of a Newtonian liquid, the analytical solution in the laminar regime is expressed as:
⃗
(4.49)
with
(
where
The error
,
and
(4.50)
correspond to the inner radius, the outer radius and the angular velocity.
between the analytical solution and the simulated velocity field can be evaluated as:
√
where
)
∑(
|⃗
⃗ |
(4.51)
)
corresponds to the total number of fluid lattice nodes and
is a
characteristic tangential velocity. Concerning the simulation with the IB-LBM, only the lattice
nodes between the inner and the outer radii are taken into account for the error computation, even
though every lattice node in the domain is considered a fluid node with this method.
The Reynolds number associated with the Couette flow problem, in the case of a Newtonian
fluid, is defined as:
(4.52)
4.2.5.1.1 Configuration
We consider a Couette geometry defined by the following parameters:
and
,
. As the flow is independent of the axial coordinates, the number of lattices in this
direction ( ) is fixed to eight in order to reduce the computational time for the finest meshes. The
total number of processors is fixed to two, the domain is split up into two subdomains along the
axis and there are thus four layers of lattices per processor.
63
To determine the impact of the three boundary condition strategies on the order of convergence
of the LBM, four different lattice spacings were used. Their values are:
[
]
(4.53)
The impact of the Reynolds number on the error is also investigated. The selected Reynolds
numbers belong to the laminar regime and their values are:
[
]
(4.54)
The corresponding angular velocities can be obtained by Eq. (4.52).
As specified in Section 4.2.3, the relaxation parameter
is set to one, so the time step
is
directly computed from the lattice spacing and the viscosity. Using Eq. (4.8), Eq. (4.10) becomes:
(4.55)
The numerical parameters associated with every boundary condition strategy were defined in
Section 4.2.4. For the EM, the fractions are computed using the exact equation of the cylinders
and the extrapolation parameters only depend on these fractions, as proposed by Guo et al. [78].
For the IB-LBM, a uniform control point distribution is used and its determination is highlighted
in Section 4.2.5.2. The boundary conditions for the top and bottom of the domain are periodic.
Finally, in order to accurately compare the errors obtained from different lattice spacings and
Reynolds numbers, a general stopping condition is needed. After conducting different tests, it
appears that the following condition is appropriate since it adapts to the flow conditions and
lattice spacing:
(4.56)
where is a given tolerance and
For the Couette flow problem,
is the characteristic length.
and
, therefore the condition reduces to:
64
(4.57)
The velocity variation
is the difference between two consecutive average velocities, which is
required to eliminate oscillations [72]:
̃
̃
(4.58)
with
(4.59)
̃
where
∑| ⃗
|
For the Couette flow problem, the average velocity
components of the velocity (i.e.
In this work, the tolerance
low value of
was fixed at
,
(4.60)
is computed using the cylindrical
and
).
for accuracy reasons. It is obvious that such a
significantly increases the computational time but, in this section, the emphasis is
put on the accuracy rather than on the computational time.
4.2.5.1.2 Results
The evolution of the error versus the lattice spacing for the three Reynolds numbers is presented
in Fig. 4.8(a) for the modified bounce-back rule (MBB), in Fig. 4.8(b) for the extrapolation
method (EM) and in Fig. 4.8(c) for the immersed boundary-lattice Boltzmann method (IB-LBM).
From Fig. 4.8(a), it appears that the MBB affects the order of convergence of the LBM by
reducing it to slightly above one. This reduction has already been observed [73, 74] and can be
explained by the inexact discretization of the cylinders. Equally interesting is that the results are
not affected by the Reynolds number in the range of values considered. The results of Fig. 4.8(b)
show that the extrapolation method is capable of conserving the second order accuracy of the
LBM, and the accuracy is hardly affected by the Reynolds number. Finally, from Fig. 4.8(c), it
65
seems that the use of the IB-LBM, as in the case of the MBB, leads to a reduction in the order of
convergence. It is important to mention that a previous study showed that the IB-LBM is second
order accurate when applied to the decaying vortex problem [83]. This difference can be
explained by the fact that the flow is not driven by the boundary conditions, as the IB-LBM was
only used to perturb the flow. Moreover, note that the IB-LBM also affects the stability of the
LBM scheme. One way to assess this stability consists of determining the maximum Mach
number (or equivalently the maximum Reynolds number, since the relaxation number is fixed to
1 in this work) that can be simulated in a stable manner with the IB-LBM as well as the other two
boundary condition strategies. Our tests have revealed that the maximum Mach numbers with the
MBB and the EM are close to 0.3. Conversely, this maximum value was reduced to 0.1 with the
IB-LBM. As a result, all simulations were divergent for
with the IB-LBM.
Figure 4.8: Convergence study for the Couette flow problem with (a) the MBB, (b) the EM and
(c) the IB-LBM
In conclusion, it appears from Fig. 4.8 that the EM is the most accurate approach and is the only
strategy capable of conserving the second order accuracy of the LBM scheme for this Couette
flow problem. The results obtained with the MBB and the IB-LBM are similar although less
accurate than those obtained with the EM. Furthermore, convergence profiles indicating firstorder accuracy are observed with these two techniques.
66
Figure 4.9: Angular velocity profiles obtained with (a) the MBB, (b) the EM and the (c) IB-LBM.
The grey line represents the analytical solution
In order to better understand the differences in accuracy, the angular velocity profiles and the
corresponding local errors are analyzed. In Fig. 4.9, the velocity profiles with the lattice spacing
are plotted for the three boundary condition strategies. From Fig. 4.9(a), it appears
that the simulated profile with the MBB oscillates around the analytical solution. In addition,
these oscillations are observed in the entire domain, even if they appear to be slightly higher near
the inner radius. For the EM (Fig. 4.9(b)), the simulated profile fits accurately with the analytical
solution. For the IB-LBM (Fig. 4.9(c)), the simulated profile globally coincides with the
analytical solution, except near the boundaries at
and
where oscillations can be
observed.
In Fig. 4.10, the average local errors on the angular velocity (i.e. the average error with respect to
the radius) with the lattice spacing
are plotted for the three boundary condition
strategies. According to Fig. 4.10(a), it appears that the error profile with the MBB is quite
sensitive to the radius, which comes from the fact that the solution is perturbed by oscillations, as
observed in Fig. 4.9(a). For the EM in Fig. 4.10(b), the error remains very small and rather
constant despite a slight increase near the inner radius. Finally, with the IB-LBM, it can be
observed in Fig. 4.10(c) that the global error is clearly dominated by errors near the inner radius.
These results signify that, with the three boundary conditions approaches, the maximum error is
situated near the boundary where a non-zero velocity is imposed.
67
Figure 4.10: Average local error profiles with (a) the MBB, (b) the EM and (c) the IB-LBM
4.2.5.2 Boundary conditions parameters
As discussed in Section 4.2.4, only the EM and the IB-LBM require numerical parameters. For
the EM, the extrapolation parameters can be seen as free parameters that are chosen to minimize
the error. Since the surface of a solid part cannot be generally described by an analytical
expression, the impact of its discretization needs to be investigated. In the same manner, the
effect of the control point distribution for the IB-LBM needs to be analyzed.
Figure 4.11: Error study on the extrapolation parameter of the EM [(F) refers to the use of the
free parameter; (C) refers to the regular definition]
68
As seen in Section 4.2.4.2, the use of the free formulation [74] for the extrapolation parameters in
the EM, for the special case
, reduces the total number of numerical parameters to one.
Thus, only the free parameter
needs to be selected. Fig. 4.11 compares the variation of the
error on velocity with respect to
(F) to the results obtained by the regular definition (C)
proposed by Guo et al. [78]. Given that the impact of the Reynolds number is negligible, the
study is limited to
. From these profiles it appears that, regardless of the lattice
spacing, only a restricted range of
, the values lower than -0.5, leads to an error inferior to that
with the regular definition. Unfortunately, this improvement in accuracy is combined with a
reduction in stability, which is amplified for fine lattice spacings. In order to conserve the
stability, the regular definition of the extrapolation parameters seems to be safer for the
simulation of industrial mixing systems. Nevertheless, this conclusion is limited to the case
, and it was shown in [74] that is possible to choose free parameters that improve significantly
the accuracy of simulation results, without affecting the stability.
Figure 4.12: Variation of the error versus the mesh spacing for the EM (the grey line refers to the
analytical fraction, the circles refer to fraction computed using a mesh)
In addition, the EM requires the computation of the fractions
that allow to accurately describe
the shape of the solid parts. As mentioned in Section 4.2.4.2, impeller surfaces cannot be
described by an analytical equation, so that meshes are used in order to do so. This means that
such a discretization is included in the boundary condition strategies, and its impact therefore
needs to be analyzed. The Couette flow problem can be seen as the worst-case scenario, because
all the boundaries are curved, meaning every fraction
will be affected by the discretization of
69
the solid parts. The variation of the error with respect to the mesh spacing is plotted in Fig. 4.12.
The lattice spacing is
and the Reynolds number is
. From these profiles,
it appears that the results converge to the one obtained with the analytical fraction. This result
signifies that the second order accuracy of the LBM can be conserved by using fine enough
obstacle mesh spacing for the solid parts.
Figure 4.13: Impact of the discretization of the inner and outer cylinders on the accuracy for the
EM (the squares refer to the analytical fraction, the circles refer to fraction computed using a
mesh)
In practice, in order to limit the total number of nodes of the surface mesh and memory usage, it
would be more convenient to link the mesh spacing to the lattice spacing. By keeping the ratio of
these two spacings constant, the second order accuracy can be conserved. For example, Fig. 4.13
shows that, for a ratio equal to one, the impact of such a strategy on the accuracy is small and the
second order accuracy is preserved.
In the case of the IB-LBM (Section 4.2.4.3), the obstacle is discretized in order to obtain the
control point distribution, and these points are then used to enforce the velocity on the solid part.
This approach requires solving a system of equations defined by Eq. (4.44), where the matrix
(Eq. (4.45)) depends on the Dirac function, the coordinates of the control points and the lattice
spacing. Given that the Dirac function has been chosen (Eq. (4.46)), the impact of the control
point distribution on the error can be analyzed. Using the term
, defined by Eq. (4.48), which
is global and related to the total number of control points and the lattice spacing, this analysis can
70
be easily done. After several tests, it appears that in order to minimize the error, the control point
distribution must be as symmetric and regular as possible. For the Couette flow problem, the
discretization was achieved as follows: the inner and outer cylinders were discretized into circles
with
and
points respectively, which correspond to a constant distance
adjacent control points (see Fig. 4.14(a)). This way, the term
total number of control point
parameter
is directly related to , as the
is proportional to . In Fig. 4.14(b), error profiles with respect to
are plotted for four different lattice spacings and a Reynolds number of
be observed that the optimal value of
related to
between any two
. It can
is located between 8 and 14. In that the term
is
and cannot be freely chosen, only a finite number of values is possible. For example
corresponds to
,
corresponds to
, and no intermediate
value exists. It is noteworthy that the error curves level off to constant values when
.
Figure 4.14: Impact of the control point distribution on the error for the IB-LBM
In order to help select the best value for
, Fig. 4.15 demonstrates its impact on the order of
convergence for the IB-LBM. The three selected values of
(corresponding to
are 6.21, 8.27 and 8.76
18, 17 and 16) and the Reynolds number is
. It clearly appears that
yields the smallest error for all lattice spacings and the best order of convergence.
This value was used for the mixing simulations carried out in this work.
71
Figure 4.15: Impact of the number of control points on the order of convergence of the IB-LBM
4.2.6 Mixing simulations
On the basis of the findings of the previous section, the boundary condition parameters can now
be chosen adequately for the simulation of fluid flow in industrial mixing systems. Two impellers
are selected, the standard anchor and the Paravisc impeller (Ekato GmbH) (see Fig. 4.16), which
are both used for the mixing of viscous fluids.
Figure 4.16: Geometry of the anchor impeller (left) and the Paravisc (Ekato GmbH) impeller
(right).
For the mixing system with the anchor, the diameter of the tank is
impeller is
diameter of the tank is
and the height of fluid is
, the diameter of the
. In the case of the Paravisc, the
, the diameter of the impeller is
and the height of
72
fluid is
, as in Iranshahi et al. [41]. The simulated fluid is a Newtonian diluted corn
syrup made up of 75% glucose. The corresponding density and viscosity are
and
. In both cases, the bottom of the tank is considered to be flat.
For a mixing system, the Reynolds number is defined as follows:
(4.61)
where
is the rotational speed (
).
4.2.6.1 Lagrangian frame of reference
As mentioned in the introduction, the Lagrangian frame of reference is used, thus the impeller is
consider to be fixed and the tank is rotating. This change of frame greatly simplifies the
simulation process as it avoids simulating the motion of the impeller and thereby prevents the
need to create new fluid lattices at each time iteration. In fact, the fluid domain is fixed in time so
the discretization process (i.e. the selection of the fluid and solid lattice nodes) only has to be
done once and the amount of data to be distributed amongst the processors is fixed. Such a
configuration yields a steady state solution. Moreover, the simulations that were performed
correspond to laminar regime with
. Under these conditions, the Coriolis force owing to
the use of a rotating frame of reference is negligible.
4.2.6.2 Dimensionless numbers
The efficiency of a mixing system is determined by its capability to agitate or mix and the power
consumption. In order to characterize the power consumption
dimensionless power number
(W) of a mixing system, the
can be used:
(4.62)
In the laminar regime, the dimensionless number
is preferred, given that its value is fixed for
a given mixing system and is only related to its geometrical parameters. It is related to the power
number
and
:
73
(4.63)
By injecting Eq. (4.62) into the previous definition, we obtain:
(4.64)
As for the power consumption, it can be computed from the entire velocity field through the
following equation [131]:
∫
where is the stress tensor (
(4.65)
̇
) and ̇ is the deformation rate tensor (
).
In the case of a Newtonian fluid, Eq. (4.65) reduces to:
∫
(4.66)
̇ ̇
where
̇
with ⃗
(
and
)
(4.67)
.
There are two ways to compute the deformation rate within the context of the LBM, directly from
the local populations or from the underlying spatial derivatives of the velocity, which in turn can
be computed using finite difference formulas. As in the present implementation of the LBM the
populations are not stored (see Section 4.2.3.1), the derivatives of the velocity are computed by a
second-order centered formula:
(
where
)
(
)
refers to the unit vector associated to the coordinate .
(4.68)
74
Near the solid parts one-sided forward and backward formulas are used:
(
)
(
)
(4.69)
In order to determine the axial pumping capacity of the mixing system, the axial flow number
is used:
(4.70)
with the axial pumping rate
given by:
∫ (∫
where
and
)
∫ (∫
(4.71)
)
correspond respectively to the positive ( when negative) and the negative (
when positive) axial velocities, implying that the velocity field must conserve mass.
For the MBB and the EM, values of
and
were computed using the velocities at the fluid
lattice nodes, as the solid lattice nodes represent the solid parts (see Section 4.2.6.4). In the case
of the IB-LBM, as every lattice in the domain is fluid, all the lattices with a radius greater than
the one of the tank were removed from the computation. Given that the volume of the impeller is
small compared to that of the tank, the impact of the corresponding fluid lattice nodes in the
computation of the dimensionless numbers is negligible.
4.2.6.3 Studied variables
In order to determine the impact of the boundary condition strategies on
and
, four different
lattice spacings were considered. For the anchor, their values are:
[
]
(4.72)
]
(4.73)
For the Paravisc, they are given by:
[
75
The impact of the Reynolds number was also investigated. The selected Reynolds numbers
belong to the laminar regime and their values are:
[
]
(4.74)
The corresponding rotational speeds, which can be obtained from Eq. (4.61), are [
] for the anchor and [
] for the
Paravisc.
The stopping condition defined by Eq. (4.56) was used, with a tolerance
fixed at
. This
value appears to be a good compromise between computational time and convergence.
4.2.6.4 Boundary conditions
As stated in Section 4.2.4, the different boundary condition strategies manage the solid parts
differently. As for the modified bounce-back rule (MBB) and the extrapolation method (EM), the
pre-processing step consists of determining which lattice nodes are solid inside the computational
domain. For the tank, this step only requires selecting the lattices whose center has a r value
larger than the radius of the tank. For the impeller, a volumetric mesh is used, so that the lattices
with centers inside one element of the mesh are considered to be solid. With the EM strategy, a
second operation is required, which computes the fraction
between the boundary lattices and
the neighboring solid part. The fractions which are used to describe the tank are computed
analytically, and the surface mesh obtained from the volumetric mesh is used to compute the
fractions which describe the impeller. Following the conclusions of Section 4.2.5.2 for the EM,
the ratio between the mesh size for the impeller and the lattice spacing was fixed to one, and the
extrapolation parameters were computed as originally proposed by Guo et al [78]. In the present
work, I-DEAS (Siemens) and Gambit (ANSYS) were used for the mesh generation of the
impellers.
Contrary to the MBB and the EM, the immersed boundary-lattice Boltzmann method (IB-LBM)
does not rely on the use of solid lattice nodes, so no search for such nodes is required. However,
surface meshes are used to generate the control points needed by this method. Following the
conclusions of Section 4.2.5.2, for every lattice spacing, the meshes were constructed by keeping
the parameter
at the optimal value (
) found for the Couette flow. It is important to
76
mention that due to the use of the Lagrangian frame of reference, the control points representing
the tank are each associated with a velocity that must converge to a value corresponding the
kinematics of this tank. To limit the artificial interaction between the tank and the boundary of
the computational domain, at least 20 lattices nodes were positioned along the gap between these
two regions. This strategy was observed to yield values of
and
that do not depend on this
gap. In the case of the Paravisc, the total number of fluid and solid lattice nodes for the three
boundary condition strategies is summarized in Table 4.1.
Tableau 4.1: Total number of fluid (F) and solid (S) lattice nodes for the Paravisc impeller
system.
Lattice spacing (mm)
MBB & EM
IB-LBM
5
2.5
1.25
0.625
F: 293,450
F: 2,350,636
F: 18,810,322
F: 150,441,376
S: 270,796
S: 1,952,250
S: 14,781,044
S: 114,987,750
F: 877,286
F: 5,438,326
F: 37,907,606
F: 282,250,966
The fluid surface on top of the tank is approximated by a free surface boundary condition [58],
which means that the surface is flat with a zero axial velocity. In the laminar regime, this
approximation has been commonly used and has shown to be valid in various situations.
4.2.6.5 Results
4.2.6.5.1 Anchor mixer
The anchor impeller (Fig. 4.16) is one of the most basic impellers, as its geometry is quite simple.
The associated flow is mainly rotational, indicating that the axial pumping is weak.
Consequently, only the power consumption is considered in this study. The variations of the
power number
with respect to the three values of the Reynolds number are plotted in Fig. 4.17
for (a) the MBB, (b) the EM and (c) the IB-LBM. From these three curves, it appears that,
regardless of the discretization and the boundary condition strategy, the slope is equal to -1,
which implies from Eq. (4.63) that
is constant.
77
Figure 4.17: Power number for the anchor impeller system: (a) the MBB, (b) the EM and (c) the
IB-LBM
In addition, it is essential to analyze the impact of the lattice spacing on the value of
. Fig. 4.18
shows its variations for the three different boundary condition strategies, at
. The
results are compared with experimental data from Foucault et al. [132] and a value obtained with
the finite element method (FEM) using software POLY3D™. These FEM simulations used
193,000 linear
tetrahedral (MINI) elements, corresponding to a total number of 118,000
velocity-pressure equations. The underlying linear systems were solved by means of a Krylovbased Uzawa solver. First, it can be noticed that the results obtained with all the boundary
conditions strategies converge to values close to experimental data and the FEM result when the
lattice spacing
is decreased. Second, the results with the MBB and the EM are close and the
convergence rate is good with variations under 2% when
goes from
. The convergence profile is also similar for the IB-LBM but the values of
to
remain less
than 10% lower. This indicates that despite the fact that convergence is observed, this boundary
condition strategy is not as accurate as the MBB and the EM. The discrepancy can be explained
by the high error in velocity near the solid parts as discussed in Section 4.2.5.1.2.
78
Figure 4.18: Convergence of Kp with respect to the lattice spacing for the anchor impeller system
4.2.6.5.2 Paravisc mixer
The Paravisc impeller (Fig. 4.16) can be considered as an improved anchor, in which the vertical
arms have been twisted to improve the axial pumping rate, in a similar way as the helical ribbon
impeller. For this mixing configuration, values of
and
obtained with the three boundary
condition strategies are compared to results available in the literature. The experimental data
come from Delaplace et al. [133] and the FEM results are from Iranshahi et al. [41]. In the latter
work, 130,000 linear
tetrahedral elements were used, corresponding to a total number of
940,000 velocity-pressure equations. The underlying linear systems were solved by means of a
Krylov-based Uzawa solver.
The variations of the power draw, at
, with respect to the lattice spacing is plotted in
Fig. 4.19. As for the anchor impeller, the results exhibit a good agreement with the experimental
data and the FEM results. Once again, the results with the MBB and the EM are close and the
convergence rate is good, as the
values between the two smallest lattice spacings differ by less
than 2%. With the IB-LBM, the convergence profile is different and, surprisingly enough, the
variation of the
values are negligible for the three smallest lattice spacings. As will be seen
next, this behavior does not hold for the axial flow number.
79
Figure 4.19: Convergence of Kp with respect to the lattice spacing for the Paravisc impeller
system
Figure 4.20: Convergence of the axial flow number
with respect to the lattice spacing for the
Paravisc impeller system
The variations of the axial flow number
with respect to the lattice spacing are plotted in Fig.
4.20. The results with the MBB and the EM are similar and agree well with the FEM results. The
influence of the lattice size is observed to be rather small with results between the two smallest
lattice spacings that differ by only 0.2%. This difference can be explained by the fact that only
the axial component of the velocity is used, which means that the error on the other components
does not appear in the axial flow number. As for the IB-LBM, the values obtained are much
lower than those of the other two approaches and the FEM. In addition, the variations of the axial
80
flow number with respect to the lattice spacing are larger, which is due to the fact that this
boundary condition strategy cannot accurately model impeller-wall interactions for such a small
wall clearance. It can be inferred from Figs. 4.19 and 4.20 that a converged value of
guarantee that the flow field is predicated accurately, which shows that
does not
is more sensitive than
to local errors on velocity. For instance, Fig. 4.21 presents contours of the norm of the
velocity on a horizontal cross-section at
, for
and
, in the case
of the three different boundary condition strategies. Figs. 4.21(a) and 4.21(b) confirm that the
flow fields obtained with the MBB and the EM are similar, as no significant differences can be
observed on the norm of the velocity. However, Fig. 4.21(c) clearly shows that the flow around
the impeller is not as accurately predicted by the IB-LBM. In fact, even if the flow is globally
well described with this approach, the contours reveal that this strategy can overestimate the local
velocity in a specific area (e.g. upper left part of the figure) while it underestimates it in other
locations (e.g lower right part of the figure). These erroneous velocities observed in the zones of
the impeller may result in streamlines going through it.
Figure 4.21: Contours of the norm of the velocity on a horizontal cross-section with (a) the MBB,
(b) the EM and the (c) IB-LBM. (
,
and
4.2.7 Parallel performance and memory usage
As specified in Section 4.2.4, the different boundary condition strategies can perturb the
intrinsically high parallel efficiency of the lattice Boltzmann method. In the case of the modified
bounce-back rule (MBB), as with the original half way bounce-back rule, the efficiency of the
LBM scheme is conserved due to the locality of the operations inherent to this boundary
81
condition strategy. Unfortunately, this is not true for the other strategies, as they rely on
interpolations or extrapolations that require non-local computations.
For the extrapolation method (EM), as previously discussed in Section 4.2.4.2, the computations
only require information from the neighboring lattices used for the propagation step.
Consequently, this strategy does not require modifications to the communication scheme between
the processors; it thus conserves the parallel efficiency of the standard LBM. As the extrapolation
computations are simple, the workload is not perturbed. The only real issue with this boundary
condition strategy is the computation of fractions for solid parts with complex shapes. In this
work, the surfaces of solid parts are described using meshes, which imply matching the boundary
lattices with the corresponding mesh elements using a search algorithm. However, since the
Lagrangian frame of reference is used, the location of these solid parts is fixed and the fractions
are computed only once as a pre-processing operation.
The immersed boundary-lattice Boltzmann method (IB-LBM) strategy is completely different
since it consists of a global approach. As a result, in order to compute the required body force, a
system of equations needs to be solved. Although the size of the system is proportional to the
total number of control points, which is small compared to the total number of lattice nodes, the
computational time associated to its solution can be important and increase the workload.
Furthermore, the system of equations needs to be solved at each time step, as the right-hand-side
is a function of the current flow field. Once again, due to the use of the Lagrangian frame of
reference, the process is simplified as the position of the control points is fixed, which avoids
recomputing the matrix at each time step. Unfortunately, the solution of the system does not only
increase the computational time, it also needs a more complex communication scheme between
the processors, which must be executed at each time step. First, the right-hand-side needs to be
built by gathering information. Second, the solution of the linear system (i.e. the body force)
needs to be distributed among the processors. Finally, the last communication step concerns the
summation of corrections on the lattice nodes near the communication zones (see Section
4.2.4.3). It is also important to mention that the solution of the system of equations is not parallel
in our implementation, so only one processor is used during this step. This leads to a slight
decrease of the parallel efficiency of the IB-LBM.
82
Tableau 4.2: Characteristics of the supercomputer used for the simulation.
Characteristics
Mammouth
Number of nodes
576
Type of node
One Intel Xeon 3.6 GHz 64-bit twincore processor
Total number of core
1152
Memory per node
8 GB
Total available memory
4.5 TB
Network
Infiniband DDR (16 Gb/s)
Peak performance
9 Tflops
To analyze the impact of the three boundary condition strategies on the workload, the
computational times associated with the simulation of the Paravisc impeller system were
evaluated. The computations were performed on a super computer (Calcul Quebec Mammouthparallel supercomputer), the characteristics of which are summarized in Table 4.2. The average
performance is on the order of 100 million lattice-site updates per second.
Figure 4.22: Graph of the computational time vs the number of processors
83
The Reynolds number was fixed at
and the tolerance
was fixed at
, the selected lattice spacing was
. The computational times associated with these
three boundary condition strategies are plotted with respect to the number of processors in Fig.
4.22. From these curves, it appears that the computational times with the MBB and the EM are
similar, which means that the cost associated with the extrapolation computations is negligible,
even if these computations are not fully local. The slope of the curve is close to -1, indicating a
linear speed up. As for the IB-LBM, the computational times are higher. This is not surprising
since this strategy needs to solve a system of equations in addition to performing the classical
collision and propagation steps. The slope of the curve is also modified, which implies that the
parallel efficiency is deteriorated by the solution of this system and supplementary inter-process
communications.
Tableau 4.3: Memory usage for the three boundary conditions strategies.
Lattice spacing (mm)
5
2.5
1.25
0.625
MBB (reference)
26.32 MB
204.3 MB
1.61 GB
12.77 GB
EM (scheme)
26.32 MB
204.3 MB
1.61 GB
12.77 GB
1.328 MB
5.418 MB
21.73 MB
85.34 MB
(+5%)
(+2.6%)
(+1.3%)
(+0.7%)
27.65 MB
209.7 MB
1.632 GB
12.85 GB
54.39 MB
337.2 MB
2.35 GB
17.5 GB
(+107%)
(+65%)
(+46%)
(+37%)
0.266 MB
0.684 MB
2.826 MB
11.29 MB
(+1%)
(+0.3%)
(+0.2%)
(+0.08%)
54.66 MB
337.9 MB
2.353 GB
17.51 GB
EM (BC)
EM (total)
IB-LBM (scheme)
IB-LBM (BC)
IB-LBM (total)
It is also important to examine the impact of the boundary condition strategies on the memory
usage. Once again, the MBB does not require any modification, so it is considered as the
reference. The EM requires the computation of fractions that need to be stored. The
supplementary memory required is then proportional to the total number of pairs of solid-fluid
84
lattice nodes. The IB-LBM is based on a completely different approach, hence the memory usage
is different. First, the system of equations and the control point coordinates need to be stored.
Given that the system matrix is sparse, the Compressed Sparse Row (CSR) format [134] is used
here to reduce memory usage. Second, because all the lattice nodes are fluid, the required
memory is increased in comparison to the MBB and the EM. This supplementary memory
corresponds to the storage of the velocity and the density information on those lattices that are
considered solid when using the MBB or the EM. More memory is also required because of the
increase of the size of the domain to reduce the interactions between the control points and its
boundary (see Section 4.2.6.4). Table 4.3 sums up the memory usage associated with the
simulation of the fluid flow in the Paravisc impeller system for four different lattice spacings. In
that the MBB is considered as the reference, the memory usage for the EM and the IB-LBM is
divided into two parts: the storage of the variables for the regular LBM scheme (referred to as
scheme) and the storage associated with the boundary condition strategy (referred to as BC).
According to this table, it seems that the supplementary memory required by the boundary
condition strategies (BC) is negligible for both the EM (from 0.7% to 5%) and the IB-LBM (from
0.08% to 1%). This result can be explained by the fact the boundary conditions are related to
surfaces, so the amount of information is negligible compared to that required for the entire
volume. Obviously, this conclusion is limited to mixing systems with a large volume to surface
ratio. The use of the IB-LBM remarkably increases the memory required for the scheme, from
37% (smallest lattice size) to 107% (largest lattice size). This is due to the increase in the size of
the domain to reduce the interactions between the control points and its boundary (see Section
4.2.6.4).
4.2.8 Conclusion
The objective of this work was to characterize the suitability of three different strategies for
imposing boundary conditions in the lattice Boltzmann method for the 3D simulation of viscous
mixing flows. The three selected strategies were the modified bounce-back (MBB) rule, the
extrapolation method (EM) and the immersed boundary-lattice Boltzmann method (IB-LBM).
Thanks to the simulation of the Couette flow problem, the impact of the three strategies on the
spatial convergence was analyzed. It was shown that only the EM is capable of conserving the
second order accuracy of the LBM while yielding the most accurate results. The deterioration
85
observed with the other strategies was also pointed out. The error with the MBB was related to
oscillations in the entire computational domain, and the error with the IB-LBM appears to be due
to a high error on the velocity close to the solid parts. The parameters of the EM and the IB-LBM
were also investigated. For the EM, it appears that the original formulation of the extrapolation
parameters was the most stable choice. Regarding the computation of the fractions, the
description of the solid part surfaces by meshes was tested and showed that, in order to conserve
the second order accuracy of the scheme, the ratio between the lattice spacing and the mesh size
must be constant. As for the IB-LBM, it was shown that an optimal value of
exists for the
distribution of control points.
The boundary condition strategies were also compared on the simulation of anchor and Paravisc
impeller systems (Section 4.2.6). Using the Lagrangian frame of reference, the accuracy of these
three boundary condition strategies was quantified by means of the power constant
axial flow number
and the
. It was shown that the LBM results are in accord with the experimental
data and finite element simulation results, and that, surprisingly enough, the MBB provides a
reliable accuracy despite its simplicity. In fact, the results and the convergence profiles of the
MBB and the EM were similar on the basis of both dimensionless numbers. With the IB-LBM,
even if the results were not so far from the experimental data and the finite element simulation
results, the predicted values of
and
and their related convergence, with respect to lattice
spacing, were different compared to those of the other two boundary condition strategies. This
indicates that the IB-LBM yields a somewhat different flow field. This was further confirmed by
comparing the contours of the norm of the velocity on a cross-section obtained with the three
different approaches.
Finally, the impact of the boundary condition strategies on the workload and memory usage was
analyzed for the Paravisc impeller system (Section 4.2.7). It appears that the MBB and the EM
both conserve the high parallel efficiency of the standard LBM. Moreover, the impact of the
extrapolation computations on the total simulation time is small. For the IB-LBM, as it is based
on a global approach and requires solving a system of equations, the computational time and the
workload are both increased, so the parallel efficiency is reduced. Lastly, even though the EM
and the IB-LBM require the storing of additional data, no significant impact was noticed as
86
regards the memory usage. In summary, this work showed that the MBB is a good compromise
for the simulation of viscous mixing flows.
4.2.9 Acknowledgments
The computer resources and support from Calcul Quebec and the financial support of the Natural
Science and Engineering Research Council of Canada (NSERC) are gratefully acknowledged.
We also would like to thank Dr. D. Vidal and Dr. F. Drolet from FPInnovations for useful
discussions.
87
CHAPITRE 5
ÉTUDE DE LA PRÉCISION DE LA LBM DANS LE CAS
DE GÉOMÉTRIES COMPLEXES
5.1 Présentation du deuxième article
Soumis à International Journal for Numerical Methods in Fluids le 23 Novembre 2012.
Auteurs : Vincent Stobiac, Philippe Tanguy, François Bertrand.
Cet article traite de l’applicabilité de la méthode d’extrapolation pour la simulation du mélange
visqueux avec la méthode de Boltzmann sur réseau. Suite à la comparaison des stratégies
d’imposition de conditions aux limites présentée dans l’article 1, il apparait que la méthode
d’extrapolation permet de traiter plus précisément les géométries complexes, et est donc mieux
adaptée pour la simulation d’écoulements dans des mélangeurs raclants. Dans cet article, la
qualité des résultats est caractérisée sur les prédictions de consommation de puissance et de
capacité de pompage au sein du système de mélange Maxblend. Ce mélangeur a été choisi car il
présente de bonnes performances sur une large gamme de nombre de Reynolds. Tout d’abord, la
sensibilité de la précision de la LBM par rapport aux conditions aux limites est étudiée sur
l’écoulement de Couette cylindrique. Il ressort qu’une petite modification des conditions aux
limites peut largement réduire la précision des prédictions de taux de cisaillement et de
consommation de puissance. Ensuite, l’écoulement dans le Maxblend est simulé à l’aide d’un
code LBM massivement parallèle. Pour la configuration sans contre-pale, les prédictions de
consommations de puissance et de création de pompage sont en accord avec les données
disponibles dans la littérature. En complément, l’impact de l’espacement entre la pale et le fond
de la cuve a aussi été étudié, montrant que la valeur standard n’est pas optimale en régime
transitoire. Finalement, pour la configuration la plus complexe, c’est-à-dire avec contre-pale, les
simulations montrent que des perturbations apparaissent autour de la pale. Au final, les résultats
cet article montrent que la méthode de Boltzmann sur réseau, couplée avec la méthode
d’extrapolation, est un outil efficace pour la simulation d’écoulements visqueux dans des
mélangeurs raclants.
88
5.2 A thorough investigation of the accuracy of the extrapolation method for
the lattice Boltzmann simulation of viscous fluid flow in a Maxblend
impeller system
5.2.1 Summary
This paper offers a thorough assessment on the performance of the extrapolation method for the
lattice Boltzmann simulation of viscous mixing flows. This method appears to be well-suited for
the treatment of the complex boundary conditions found in various mixing systems. Here, the
ability to simulate accurate power consumption and pumping capacity is evaluated on several
configurations of the Maxblend mixing system, which has proven efficient in a wide range of
applications. First, the impact of the boundary conditions on the spatial convergence of the lattice
Boltzmann method (LBM) is determined on the 3D Couette flow, clearly showing that small
modifications of the boundary conditions may reduce the accuracy of the predicted shear rate and
power. Second, a parallel LBM scheme was used to simulate fluid flow within a Maxblend
mixing system. For the unbaffled configuration, the simulated power consumption and the
pumping capacity are observed to be in good agreement with experimental data and finite
element simulation results. The effect of the bottom clearance is also successfully evaluated,
suggesting that the standard bottom clearance is not optimum in the transitional regime. Lastly,
results for the most geometrically complex case (baffled configuration) indicate that the power
consumption is affected by numerical perturbations appearing around the moving impeller.
Overall, these results show that, when combined with the extrapolation method for the treatment
of boundary conditions, the LBM is an efficient tool for the investigation of viscous flow in
mixers of industrial relevance.
5.2.2 Introduction
The lattice Boltzmann method (LBM) is a powerful numerical method that has been successfully
applied for the simulation of various industrial flows, such as those in catalytic converters [7,
135], stirred tank reactors [27, 98] and static mixers [136, 137]. This method differs from the
classical CFD methods, the finite element method (FEM) and the finite volume method (FVM),
in that the solution obtained for the Navier-Stokes equations uses the mesoscopic scale rather
than the macroscopic scale. Owing to this change of scale, the modeling of several hydrodynamic
89
phenomena, such as those induced by turbulence [86, 116, 138] or complex rheology [16, 54,
115], is simplified. The LBM also has a linear constitutive equation (the lattice Boltzmann
equation) both simple and local. When applied on a structured Cartesian grid, such advantages
give the LBM its highly intrinsic parallel performance, which has been favourably used to reduce
simulation costs for geometrically complex flow problems involving, for instance, porous media
[3, 110, 115] or solid-fluid suspensions [72, 111, 139].
Inversely, the major drawback of the LBM is in the definition of macroscopic boundary
conditions, which are crucial for complex geometries. While the so-called modified bounce-back
rule [72] has been commonly considered to impose a non-zero velocity along curved boundaries,
its use is known to decrease the accuracy of the LBM scheme [73, 74, 140]. Lately, alternate
strategies have been proposed to reduce these inaccuracies, namely the extrapolation or
interpolation methods [76-79] and the immersed boundary methods [16, 81, 83].
Simulation of mixing in stirred tanks is a challenging problem of industrial relevance, the results
of which can serve for process design and optimization. The geometries of stirred tanks are
highly complex and present several difficulties, such as the motion of the impeller or the small
gap between this impeller and the tank wall. In this context, the LBM has been successfully used
for the simulation of turbulent mixing flows [27, 54, 86, 117, 141, 142]. In mixing systems
comprised of a Rushton turbine, such simulations have permitted to investigate flow fields in the
vicinity of the impeller [27, 54, 86, 117, 141, 142], as well as evaluate the turbulent kinetic
energy [54, 86, 117, 141, 142], the energy dissipation rate [86, 117], the apparent viscosity [54]
and the volume of caverns in the case of Bingham fluids [27].
In contrast to the mixing occurring inside the vortices of turbulent flows, the mixing in viscous
flow systems is driven by the shear-induced stretching, folding and breaking mechanisms of the
fluid streams generated by the close-clearance impeller. The complexity of the geometry is
therefore increased by the presence of a small wall clearance, within which it becomes essential
to accurately capture the flow. We believe the difficulty in imposing accurate boundary
conditions is one of the reasons why the LBM has rarely been used in this context. In our
previous study [140], we showed that, in the case of a fixed geometry, the different boundary
condition strategies investigated, let alone the modified bounce-back [72], the extrapolation [78]
and the immersed boundary [83] methods, are not all equivalent in terms of accuracy. In
90
particular, the simulation of a Couette flow showed that the extrapolation method is the only
strategy that preserves the second-order accuracy of the LBM. When applied to mixing systems,
the modified bounce-back rule and the extrapolation method (EM) were observed to be
potentially good choices for the prediction of power consumption and pumping capacity in the
case of anchor and Paravisc impeller systems. For these reasons and considering that the slightly
larger cost of the extrapolation method with respect to the modified bounce-back rule is
compensated by the gain in accuracy, the former was used in this work.
This investigation provides a thorough assessment of the accuracy of the LBM and EM
combination for the simulation of geometrically complex mixing flows. The emphasis is put on
the simulated power consumption and pumping capacity, which are two important characteristic
numbers for mixer comparison. The Maxblend impeller (Sumitomo Heavy Industries, Japan) has
been selected for this study, as it has been observed to be an efficient mixer in various
applications [35]. Several experimental and numerical investigations on design parameters [13,
33, 36, 143] are available in the literature and will be used to assess our results. We have chosen
to focus on the impact on power consumption and pumping capacity of a small variation of the
bottom clearance. The procedure for simulating the motion of the impeller is also investigated.
To our knowledge, no extended analysis for such complex 3D flows have yet been reported.
Moreover, the conclusions of the present study are not limited to mixing flows and may be
applicable to other types of geometrically complex flow situations.
The organization of the paper is as follows. In Section 5.2.3, the LBM is revisited including the
specifications of the implementation used in this work. In Section 5.2.4, the three-dimensional
Couette flow problem is simulated to verify the spatial convergence of the scheme as regards the
shear rate and the power. The impacts of the boundary conditions are also investigated. In Section
5.2.5, the lattice Boltzmann scheme is used to simulate the hydrodynamics of the Maxblend
impeller. First, using a Lagrangian frame of reference, the performance of the unbaffled mixing
system is characterized for several values of the bottom clearance. Second, the accuracy of the
predicted power consumption is studied for the baffled configuration. Concluding remarks are
given in Section 5.2.6.
91
5.2.3 The lattice Boltzmann method
Only a brief presentation of the method is provided here and the reader is referred to [58, 144] for
detailed explanations. Contrary to the classical CFD tools, the LBM is based on the kinetic theory
of gas, situated at the mesoscopic scale. The solution of the Navier-Stokes equations is then
recovered from the solution of the Boltzmann equation using populations of particles. These
populations propagate and collide along the bonds of a lattice grid and their motions are governed
by a simple recurrence equation. The main advantages of the resulting scheme are the linearity
and simplicity of the constitutive equation, giving the LBM its high parallel performance.
First we consider the population
at position , with velocity
lattice, macroscopic variables such as the density
, at time t, so that in every
and the velocity ⃗ can be easily recovered as
[65]:
(5.1)
∑
(5.2)
⃗
where
∑
is the total number of populations per lattice,
force density and
is the mean density,
is a body
is the time step. In the present work, the 15-population D3Q15 lattice [63] is
chosen since preliminary results demonstrated that, compared to lattices with more populations
(19 and 27), the accuracy is comparable and the computational time is reduced.
The lattice Boltzmann equation, coupled with the BGK collision operator [60, 63], is defined as
follows:
[
where the forcing term
]
(5.3)
is given by [65]
(
)
[
⃗
⃗
]
(5.4)
92
and the equilibrium population is defined as
[
Where
is the lattice spacing,
and the weights
(
⃗
⃗
⃗ ⃗
is the relaxation parameter,
are, for the D3Q15 lattice, if
, if
The relaxation parameter is related to the fluid viscosity
)]
(5.5)
√
and
is the speed of sound,
if
.
by the following relation:
(5.6)
The solution of the Navier-Stokes equations is recovered under the hypotheses of small Mach
(
‖⃗ ‖
) and small Knudsen (
case of diffusive scaling (i.e.
) numbers, where L is a characteristic length. In the
), the resulting explicit scheme has been proven to be
second order accurate in space [121, 144, 145] in the case of a periodic computational domain.
However, the accuracy of the scheme can be largely affected by the boundary conditions.
5.2.3.1 Boundary conditions
It is noteworthy that the LBM scheme defined by Eq. (5.3) is only valid for a lattice node
situated inside the computational domain. As a result, it must be completed by boundary
conditions at the boundary of the domain. As mentioned in the introduction, the fact that the
primary variable
is mesoscopic is a drawback as regards the imposition of macroscopic
boundary conditions on velocity. However, in the case of an exact discretization (i.e. all
boundaries are situated precisely on lattice nodes or half a grid spacing away), several simple
rules may be applied to missing populations in order to model boundary conditions. For example,
Dirichlet boundary conditions on velocity can be defined by the so-called modified bounce-back
rule [72]. In the same manner, free slip boundary conditions [58] may be used to model surfaces
without friction. These boundary condition strategies can conserve the second order accuracy of
the LBM. However, in the case of curved boundaries, such strategies are known to reduce the
accuracy to the first order [73, 74].
93
For the purposes of this work, in order to accurately describe the geometry of the mixing system,
all boundary conditions except those at the free surface are imposed with the extrapolation
method (EM) [78]. The second order accuracy is then conserved [74, 78, 140] and the exact
position of the boundaries is computed and stored. To do so, this boundary condition strategy
requires the computation of the fraction
representing the distance between the fluid node and
the boundary (see Fig. 5.1). The fraction is computed as
|
|
|
|
respectively to the position of the solid and fluid nodes, and
actual boundary and the line joining
and
, where
correspond
is the intersection between the
. Based on the decomposition of the populations
into equilibrium and non-equilibrium parts, i.e.
the fraction
and
, the EM uses
to extrapolate the velocity and the non-equilibrium populations at the solid node
.
Figure 5.1: Distance to the boundary with the extrapolation method (EM).
The unknown populations at the first fluid node
(
)
are computed as follows:
(
)
The equilibrium part is computed with Eq. (5.5), where the density
(
), and the velocity ⃗
⃗
(5.7)
is approximated by
at the solid node is determined by the following
extrapolation procedure:
⃗
⃗
⃗
⃗
(5.8)
94
where ⃗ , ⃗ and ⃗
correspond to the velocity at nodes
5.1). The formulas for the extrapolation parameters
,
and
, respectively (see Fig.
are given in Table 5.1.
Tableau 5.1: Formulas for the parameters of the extrapolation method (EM).
0
1
Next, the non-equilibrium part of Eq. (5.7) is directly extrapolated from the neighboring nodes as
(
where the extrapolation parameter
)
(
(5.9)
)
is defined in Table 5.1.
It is important to mention that Eqs. (5.8) and (5.9), which describe the boundary condition
is fluid. Otherwise, we propose to impose ⃗
strategy, are only valid if the lattice node
and
(
⃗
).
5.2.3.2 Power computation
Independent of the numerical method used, the power may be obtained in two distinct manners.
First, it can be computed as a volume integral of the viscous dissipation,
∫
̇
, where
and ̇ are the stress tensor and the shear rate tensor,
denotes the computational domain and
respectively. In the case of a Newtonian fluid, this formula reduces to
∫
̇ ̇
(5.10)
With the LBM, the shear rate tensor can be computed from the velocity field with a second-order
centered difference formula, but it can also be computed directly from the populations inside a
lattice. The latter computation is local and expressed for the
component of ̇ by [62, 89]:
95
̇
where
∑(
)
is the dimension of the problem (3 in this work) and
(5.11)
is the Kronecker delta function.
In this work, the shear rate is computed from the populations in all cases since our tests showed
that the difference between the two approaches is negligible.
Second, the power can be computed as a surface integral of the force moment acting at the
boundary, the torque, times the angular velocity :
(5.12)
∫
where
is the center of gravity,
is the boundary of the domain and
is the force per unit
surface acting at the boundary. With the LBM, this force can be readily computed from the
populations with the so-called momentum exchange algorithm [72]. More precisely, for every
direction
involving a boundary condition, the force can be approximated by [72, 89, 146]
(
where
and
)
(5.13)
refers to the post-collision time (after a collision but before the next propagation step)
is the direction opposite of .
The important fact here is that the simulation itself is not affected by the approach used to
compute the power. In other words, this computation is similar to a post-processing step.
Therefore, two different values of the power can be obtained. It can already be inferred that, since
the evaluation of
in Eq. (5.13) is less expensive than that of ̇ in Eq. (5.11), and can be done
during the computation of the boundary conditions, the method based on the torque requires less
computational time than that based on the shear rate. It is then important to determine if this
faster approach is as accurate as the other one. This will be done in the next sections.
96
5.2.4 Analysis of the spatial convergence
To verify the order of the spatial convergence on a problem with an analytical solution, the threedimensional cylindrical Couette flow is simulated (see Fig. 5.2). In this case, the fluid is
Newtonian and only the inner cylinder is rotating. In the laminar regime, the stationary analytical
solutions for the velocity, the norm of the shear rate and the power
(
‖ ̇‖
are expressed as
)
(5.14)
(5.15)
√
(5.16)
where
is the radial position,
and
are the radii of the inner and outer cylinders, and
is
the height of the fluid.
Figure 5.2: Geometry of the Couette flow problem.
The Reynolds number is set to one and the emphasis is put on the impact of the lattice spacing on
the accuracy of the scheme. The boundary conditions for the inner and outer cylinders are defined
by the extrapolation method. As presented in Section 5.2.3.2, the two different approaches to
compute the power (using the shear rate or the torque) are compared. The deviation from the
analytical solution is computed as
√ ∑
, where
stands for the total number of
97
fluid lattice nodes and the symbol “*” denotes the analytical solution of the quantity
under
investigation.
5.2.4.1 Circular Couette flow with periodic boundary conditions
In this case, the boundary conditions for the top and bottom are set as periodic. The velocities of
the inner and outer cylinders are imposed by means of the extrapolation method. The accuracy of
the LBM scheme is determined for the velocity, shear rate and power. The errors on the velocity
and norm of the shear rate are presented in Fig. 5.3(a), so that the order of the spatial convergence
can be quantified as the slope of the straight line. This figure indicates that the scheme achieves
second-order accuracy for the velocity (slope of 2.11), thereby verifying the adequacy of the
implementation, as already discussed in our previous paper [140]. It also appears that the scheme
is second order for the norm of the shear rate (slope of 1.84); these findings support what has
already been shown in the literature [62]. In the same manner, the accuracy of the LBM scheme
on the power is presented in Fig. 5.3(b). For the power computed from the torque, the scheme
achieves second-order accuracy (slope of 2.07). Conversely, when the power is computed from
the shear rate, the error does not monotonically decrease with the lattice spacing. Indeed, here the
error on the power is dominated by the error on the volume of the fluid zone. More precisely, due
to the use of a Cartesian structured mesh, the error on this volume is globally convergent but can
vary locally (not shown here). Given that the power computed from the shear rate is based on a
volume integral (see Eq. (5.10)), the error on the power can likewise be locally affected,
notwithstanding the global convergence. From the results shown in Fig. 5.3(b), it seems that
computing the power through the surface integral in Eq. (5.12) does not lead to local oscillations
in the corresponding convergence curve.
98
Figure 5.3: Convergence curves for the Couette problem with periodic boundary conditions: (a)
velocity and shear rate, (b) power consumption from shear rate and torque values. The numbers
correspond to the slopes of the straight lines.
5.2.4.2 Circular Couette flow with mixed boundary conditions
Here the boundary conditions at the top and bottom are modified to better approximate real
mixing systems. To do so, free slip boundary conditions (i.e.
and
are free and
) are
imposed at the top, mimicking a free surface. In addition, the analytical velocity defined by Eq.
(5.14) is applied at the bottom with the extrapolation method. By means of such modifications
(referred to as mixed boundary conditions), the impacts of the boundary conditions can be
studied without modifying the stationary solution of the flow.
The errors on the velocity and norm of the shear rate are presented in Fig. 5.4(a). It appears that
the scheme still achieves second-order accuracy for the velocity (slope of 2.13), thereby
indicating that the modification of the boundary conditions does not globally modify the
convergence of the computed velocity field. Conversely, the order of convergence is reduced to
one for the norm of the shear rate (slope of 1.11), as opposed to two when periodic boundary
conditions are used (Section 5.2.4.1). Fig. 5.4(b) shows that, once again, when the power is
computed from the shear rate, it doesn’t seem to be convergent (at least for
). In
addition, Fig. 5.4(b) also shows that the convergence is likewise reduced to the first order for the
power computed from the torque (slope of 1.13), as opposed to two when periodic boundary
conditions are used (Section 5.2.4.1). A similar reduction to the first order of accuracy for the
99
total force was also observed by Caiazzo et al. [146]. We believe these reductions on the
convergence can be explained by local perturbations on the solution generated by the boundary
conditions. For instance, an axial component of the velocity is created at the intersection between
the cylinders and the bottom of the Couette geometry, where the extrapolation method is illdefined (i.e. no second fluid lattice, see Section 5.2.3.1). Fig. 5.4(c) presents the values of the
sum over the lattice nodes of the axial component of the velocity, denoted
, versus the lattice
spacing. Contrary to the periodic case, a small axial flow is indeed created, but fortunately, its
value reduces as the square of the lattice spacing.
Figure 5.4: Convergence curves for the Couette flow problem with mixed boundary conditions:
(a) velocity and shear rate, (b) power consumption and (c) sum over the lattice nodes of the axial
component of the velocity,
. The numbers correspond to the slopes of the straight lines.
In summary, we observe that a modification on boundary conditions can reduce the accuracy of
the shear rate and the power, even if the scheme still achieves second-order accuracy for the
velocity. These developments are important for simulations in more complex geometries, in that
they suggest the convergence of the shear rate and the power will in all likelihood be reduced to
first-order accuracy. Moreover, the power computation based on the torque (i.e. on the force) is
seemingly more adequate owing to its monotonically decreasing convergence behavior. A similar
conclusion was drawn by Mei et al. [89] for the computation of drag coefficients over cylinders
and spheres, where the computation based on the force was found to be more accurate than the
stress integration over the surface of the obstacle. Finally, it appears that a numerical axial flow
can be generated by local perturbations on the velocity field, when the extrapolation method is
ill-defined. This result must be kept in mind as the axial flow is used to characterize the pumping
capacity of a mixing system.
100
5.2.5 Simulation results for the Maxblend mixer
A parallel implementation of the LBM presented in Section 5.2.3 was used to simulate fluid flow
within a Maxblend mixing system. First, for the unbaffled configuration, the accuracy of the
simulation process is characterized, as well as the impacts of the bottom clearance on the
pumping capacity and power consumption. Following these results, the analysis is extended to
determine the optimum values of the bottom clearance for different Reynolds numbers. Next, the
accuracy of the power consumption is investigated in the case of the baffled configuration. This
case is more complex in terms of the geometry and simulation process, in that it requires
simulating the motion of the impeller with respect to the baffles.
5.2.5.1 Methodology
The Maxblend mixer considered in this work is the wedge impeller configuration, as presented in
Fig. 5.5(a). It can be used with or without baffles, depending on the flow regime. The geometrical
characteristics of the mixing system and the physical properties of the fluid are provided in Table
5.2. As an illustration, the velocity profile at
in the unbaffled configuration is given in
Fig. 5.5(b).
(a)
(b)
Figure 5.5 : The Maxblend impeller (wedge configuration): (a) geometry and (b) illustration of
the velocity profile at Re=120.
101
In the case of a mixing system, the Reynolds number is defined as
rotational speed (in RPS) and
, where
is the
is the impeller diameter. The variation of the Reynolds number is
achieved by adjusting the rotational speed. For impeller comparison purposes [35, 36, 143], the
power consumption
as
is generally represented by the dimensionless power number
, defined
. As it is constant is the turbulent regime, it can be used to compare the
efficiency of mixing systems designed for low viscosity fluids. Conversely, the power constant
, is constant in the laminar regime. It can then be used for the
comparison of mixing systems designed for high viscosity fluids.
Tableau 5.2: Geometrical characteristics of the Maxblend mixing system and physical properties
of the fluid.
Standard bottom clearance
5.5 mm
D
Impeller diameter
255 mm
H
Fluid height
410 mm
W
Baffle width
26 mm
T
Tank diameter
356 mm
V
Vessel capacity
35.4 L
Fluid density
1240
Fluid viscosity
0.1 Pa.s
The pumping capacity of the mixing system is also represented by a dimensionless number, the
axial flow number
, where the axial pumping
is given by
∫
In this expression, H is the fluid height, V is the total fluid volume and
component of the velocity [13].
(5.17)
is the upward axial
102
In this work, two types of simulation strategies were used depending on the topology of the flow
domain. To begin with, if the topology is fixed (i.e. the shape of the fluid domain is fixed in
time), the simulation process can be simplified by the use of a Lagrangian frame of reference. For
instance, the simulation of the unbaffled mixing system can be done with a frame of reference
attached to the impeller. Owing to this change of frame, there is no need to simulate the motion
of the impeller and the discretization process (selection of solid and fluid lattice nodes) is done
only once. Because such a frame of reference is in acceleration, a body force
is used to
add the Coriolis and centrifugal forces:
⃗
⃗
where ⃗ is the rotational velocity vector (in
⃗
⃗
(5.18)
) and
is the radius. However, if the
topology of the flow domain is time dependent, as with the baffled configuration, the Lagrangian
frame of reference can no longer be used. The simulation strategy is then complexified by the
need to simulate the motion of the impeller. This modification implies re-doing the discretization
process (i.e. selection of fluid and solid lattice nodes) at each time step and initializing the new
fluid lattice nodes. To do so, we have chosen to use the so called non-equilibrium corrected refill
[88] in order to minimize oscillations around the impeller. We will come back to this point in
Section 5.2.5.3.
As mentioned in Section 5.2.3.1, all the boundary conditions, except those at the free surface, are
imposed with the extrapolation method. This boundary condition strategy requires computing the
fractions
during the discretization process. The fractions used to describe the tank walls are
computed analytically, while a surface mesh is used to compute the fractions for the impeller. At
the top of the tank, the free surface of the fluid is modeled by a free slip boundary condition (i.e.
and
are free and
). This approximation has been commonly used in laminar and
transitional regimes when the Froude number
is small.
The simulations were performed on Calcul Quebec Colosse and Mammouth-Parallèle II
supercomputers, the characteristics of which are summarized in Table 5.3.
103
Tableau 5.3: Characteristics of the supercomputers used for the simulations.
Characteristics
Colosse supercomputer
Mammouth-Parallèle II
Number of nodes
960
1588
Type of nodes
2 Intel Nehalem-Ep 4-core proc.
2 AMD Opteron 12-core proc.
Total number of cores
7680
38112
Memory per node
24 GB
32 GB
Total available memory
23 TB
50.8 TB
Network
Infiniband QDR (40 Gb/s)
Infiniband QDR (4 Gb/s)
Peak performance
88 Tflops
333 Tflops
5.2.5.2 Effect of bottom clearance
This section is primarily concerned with the suitability of the LBM to model a small variation of
the bottom clearance. The Lagrangian frame of reference was used to simulate the stationary flow
of the unbaffled mixing configuration. The simulation was stopped when the relative variation of
the flow field was lower than
. The selected lattice spacing
is equal to 1.25 mm (yielding
more than 19 million fluid lattice nodes), as preliminary results showed that, compared to finer
lattice spacings, the variations on the power constant and axial flow number were lower than 2%
and 1%, respectively. The three selected values of the Reynolds number (7, 33 and 68) are
representative of typical conditions of viscous mixing in laminar and transitional regimes. The
relaxation parameter
of the lattice Boltzmann equation (see Eq. (5.6)), was adjusted to ensure
low Mach numbers, thereby limiting compressibility errors. However, this selection leads to a
compromise between accuracy and computational time, as the simulation time drastically
increases when
. Thus, the maximum Mach numbers were
for
and
for the two larger Reynolds numbers (33 and 68), which are small enough to ensure good
accuracy (see Section 5.2.3). The computations were performed on Colosse supercomputer for an
average computational time of 7 hours with 64 cores.
104
5.2.5.2.1 Standard bottom clearance
First, it is essential to ensure the accuracy of the simulation strategy. To accomplish this, our
results are compared with experimental data from Fradette et al. [143] and simulation results
obtained with the finite element method (FEM) by Devals et al. [13]. The values of the
corresponding power constant for the standard bottom clearance
, referred to as
, are summarized in Table 5.4. Surprisingly enough, with the LBM, the power
constants computed from the shear rate (
) and torque (
) are quite similar. The
difference is rather small considering that the computation based on the shear rate was shown to
be not monotonically convergent in Section 5.2.4. These results exhibit a good agreement with
the experimental data (
), corresponding to a maximum difference of 4%, which is
within the experimental error (of the order of
). On the other hand, with the FEM (
), the difference is more significant (+8%), but can be explained by differences in design
parameters. For instance, the ratio between the diameters of the tank and impeller is 5% lower in
the FEM simulations.
Tableau 5.4: Values of the power constant for the standard bottom clearance obtained with the
LBM, with the FEM [13] and measured experimentally [143].
Method
LBM
FEM
Experimental data
177 (+8%)
164
158 from shear rate (-4%)
160 from torque (-2%)
5.2.5.2.2 Impact of the bottom clearance
Knowing the accuracy of the LBM on
, we can now focus on the impact of the bottom
clearance on the power consumption. Fig. 5.6(a) shows the variation of the normalized power
constant
versus the normalized bottom clearance
, for results obtained with the
LBM and the FEM as well as experimental data. Notice that the curves computed from the torque
and the shear rate are almost identical. Furthermore, a good agreement is observed between the
LBM and FEM results, since the sole significant differences are obtained for the smallest bottom
105
clearance, where numerical errors are higher (fewer elements between the impeller and the tank
wall). A rather good agreement is also observed with the experimental data. The differences can
be explained by the experimental error on the bottom clearance (
, i.e. 20% of
). This
rather large experimental error is due to the technical difficulty of accurately positioning the
impeller, as it is directly attached to the motor.
Figure 5.6: Comparison of LBM, FEM [13] and experimental data [143] for (a) the impact of the
bottom clearance on the power constant and (b) the impact of the Reynolds number on the axial
flow number.
It is also important to verify the accuracy of the LBM on the pumping capacity. This is crucial
because it was shown in Section 5.2.4.2 that the boundary conditions used for the LBM can
generate an erroneous numerical pumping. Fig. 5.6(b) displays the evolution of the axial flow
number
versus
for results obtained with the LBM and the FEM. A good agreement is
observed, as the differences are equal to 7.9%, 2.9% and 6.5% for the Reynolds numbers of 7, 33
and 68, respectively. Moreover, it shows that the LBM and FEM exhibit the same linear behavior
for
.
The results above show that the power consumption and pumping capacity simulated with the
LBM, combined with the extrapolation method, are consistent with experimental data and FEM
simulation results. Furthermore, the method used to compute the power (from the shear rate or
the torque) does not significantly impact the accuracy of the power consumption itself or its
dependency on the bottom clearance. Note that all power values reported in the remainder of this
106
paper were calculated form the torque. Finally, these results indicate that the erroneous pumping
inherent to the boundary conditions and the extrapolation method, as discussed in Section 5.2.4.2,
is small and does not significantly affect the pumping behavior predicted by the LBM.
5.2.5.2.3 Optimum bottom clearance
We showed in the previous section that the LBM can predict accurately the impact of a variation
of the geometry on the power consumption and the pumping capacity. For the sake of
completeness, the simulation work was extended to determine an optimum bottom clearance in
regard to these two macroscopic quantities, and show by the same token that the LBM can be
used for design and optimization purposes.
First, the impact of the bottom clearance on the power consumption is considered. Fig. 5.7(a)
compares the variation of the normalized power number
versus
, for three
Reynolds numbers (7, 33 and 68), where the power number at the standard bottom clearance is
referred to as
. According to these curves, the impact on the power increases
inversely with the bottom clearance. In addition, for normalized bottom clearance values lower
than 1.5, the variation is independent of the Reynolds number. Conversely, for
, the
impact on the power consumption is lower for the largest value of the Reynolds number (
). One possible explanation for the different behavior observed at
is the likely
increasing impact of inertia on the hydrodynamics when the clearance increases. In the same
manner, Fig. 5.7(b) shows the variation of the normalized pumping number
with
respect to the bottom clearance, where the axial flow number at the standard bottom clearance is
referred to as
. It can be observed that the profile is quasi-linear and the impact of
the bottom clearance is inversely proportional to the Reynolds number. Lastly, considering that
the pumping plays an important role in the homogeneity of a mixture and that the power
consumption represents its cost, it is worthwhile to investigate the ratio
, as was done in
the experimental work of Guntzburger et al. [35]. We believe that this ratio, which characterizes
the efficiency of the mixing process, may be used to determine an optimum bottom clearance.
From Fig. 5.7(c), it can be noticed that, in the laminar regime (
), the optimal value of the
normalized bottom clearance is between 0.8 and 1.4, therefore confirming the standard bottom
clearance (
) is a good choice. However, in the transitional regime (
and 68), the
107
optimum value is situated between 1.4 and 2, which suggests that the standard bottom clearance
is not adequately large enough. It is noteworthy that, compared to the laminar regime, the impact
of the bottom clearance is globally reduced in the transitional regime. We believe these results
are highly significant for the minimization of the power consumption, given that the standard
bottom clearance is one of the most relevant design parameters of the Maxblend [143]. It also
confirms that the LBM is useful to investigate small variations of design parameters.
Figure 5.7: Impact of the bottom clearance on (a) the power number, (b) the axial flow number
and (c) their ratio, for Reynolds numbers of 7, 33 and 68.
5.2.5.3 Simulation of rotating impeller
The objective is now to characterize the accuracy of the simulation process in the case of the
baffled configuration. The emphasis is put on the impact of the lattice spacing on the power
consumption and on the comparison of the two different approaches to compute the power (i.e.
from the torque and shear rate). Since the Lagrangian frame of reference can no longer be used,
the position of the solid part (i.e. the impeller) is not fixed in time. Therefore, the discretization
process (i.e. selection of fluid and solid lattice nodes) must be applied at each time step, thereby
possibly increasing significantly the total computational time. The motion of the solid part also
requires initializing the new fluid nodes. Unfortunately, given their mesoscopic nature, there is no
analytic definition of the populations
for a given velocity, hence no analytical initialization (or
refill) exists. Given that simple refill methods using the equilibrium population [87] generally
generate inaccuracies and oscillations around moving obstacles, we have chosen to use a more
elaborate approach, the so-called non-equilibrium corrected refill [88]. More precisely, as in the
case of the extrapolation method (see Section 5.2.3.1), it is based on the decomposition of the
108
populations into equilibrium and non-equilibrium parts to determine unknown populations.
Similarly, the equilibrium part is computed by means of Eq. (5.5) with an extrapolated velocity.
The non-equilibrium part is directly copied from that at a neighboring fluid node. In our case, the
direction used to extrapolate the velocity and select the neighboring fluid node is that
countercurrent to the flow. On account of this approach, oscillations around moving obstacles are
minimized and the accuracy of the force computation is enhanced [88].
Three selected lattice spacings were considered:
. The Reynolds numbers are
,
and
and 120, for which experimental data [143] and
FEM results [13] are available. As mentioned in the previous section, the relaxation parameter
was adjusted for the Mach number, resulting in a maximum Mach number lower than
The selected bottom clearance is
.
in order to limit inaccuracies at the bottom of the
mixing system. As the lattice spacings are small compared to the distance covered by the impeller
during one time step
, the discretization process (see Section 5.2.5.1) was only applied every 5
iterations in order to reduce the computational cost. Note that this choice was observed to yield
adequate results. The total simulation time corresponds to 3 and 10 complete revolutions of the
impeller for the Reynolds numbers of 7 and 120, respectively. All computations were performed
on Mammouth-Parallèle II supercomputer. For the 3 impeller revolutions, the average
computational time was 10 hours on 72 cores for
on 120 cores for
, 16 hours on 96 cores for
and 28 hours
. These rather large computational times can be explained by the non-
stationary nature of the solutions, which require re-doing the discretization process and
computing the power regularly throughout the simulation. As mentioned earlier, these two
procedures are time consuming and may constitute a significant portion of the total
computational time (up to 60 % in our tests).
In order to determine the accuracy of the results, the unbaffled configuration was simulated at
. In this case, the stationary result obtained in the previous section (i.e. with the
Lagrangian frame of reference) is considered to be the most accurate result and is referred to as
. For instance, Fig. 5.8 denotes the evolution of the power number for the largest lattice
spacing
. Apparently, contrary to what was observed in Section 5.2.5.2, the method used to
compute the power greatly impacts its variation in time, in that the power computed from the
torque is largely more affected by oscillations. This difference can be explained by the fact that
109
the torque is computed as a surface integral of the force on the impeller (see Eq. (5.11)), where
the refill of the new fluid nodes generates oscillations. Conversely, when the power is computed
as a volume integral of the shear rate (see Eq. (5.10)), the number of lattice nodes affected by the
refill is negligible compared to the total number of fluid nodes, thereby yielding a better
accuracy.
Figure 5.8: Evolution of the power number for the unbaffled configuration at Re=7 (
.
In order to investigate the impact of the lattice spacing, the variations of the average power
number and the standard deviations of the power number are presented in Fig. 5.9. These data
were computed over the last 2 revolutions when a quasi-steady state behavior was achieved. Fig.
5.9(a) shows that the results for the power number exhibit a good agreement with the reference
solution
, with a maximum difference of 4.5%. More precisely, while the power is
always overestimated by the two different approaches, the computation based on the shear rate
yields the most accurate average value (maximum difference of 2.4%). In Fig 5.9(b), the standard
deviation around the average power is 4 times higher with the torque computation, although it
reduces for both approaches when the lattice spacing is reduced. To sum up, in the case of a
moving solid part, the power computation based on the shear rate is seemingly more effective in
that the average power value is more accurate and the oscillations are reduced. It is also important
to underline that these oscillations are solely due to the initialization process of the new fluid
nodes, considering that they did not appear when the Lagrangian frame of reference was used.
110
Figure 5.9: Variations with respect to the lattice spacing of (a) the average power number and (b)
the standard deviation of the power number, for the unbaffled configuration at Re=7.
The results for the baffled and unbaffled configurations and a Reynolds number of 120 are next
compared. Fig. 5.10 shows the evolution of the power during the last two revolutions for the
largest lattice spacing
. First, Fig. 5.10(a) presents the results for the unbaffled configuration.
It can be noticed that, similarly to the results obtained at
(Fig. 5.8), the power computed
from the torque is largely more affected by oscillations. In addition, Fig. 5.10(b) shows that, for
the baffled configuration, the evolution of the power is periodic with respect to the relative
position of the impeller versus the baffles. A comparable evolution was observed with FEM
results by Devals et al. [13], therefore they will be used to assess our LBM results. To this end,
two mixing system parameters are considered: the increase in power consumption due to the
baffles and the maximum variation of power during one impeller revolution.
Fig. 5.11(a) compares the increase in power obtained with the LBM (this work) and the FEM
[13], as well as experimental data [143]. A good agreement is observed between the numerical
results, as the LBM yields to an increase between 11.8 and 13.1%, and the FEM gives +12%.
However, compared to experimental data (+8.5%), the impact of the baffles appears to be
overestimated. In addition, Fig. 5.11(b) compares the maximum power amplitude by the LBM
and FEM. Here, the power amplitude refers to the maximum amplitude of the power due to the
motion of the impeller with respect to the baffles. In order to vanish the oscillations related to the
refill procedure within the LBM, the variation of the power was smoothed with a moving average
111
over 20 iterations. According to these curves, it appears that only the amplitude obtained with the
power computed from the shear rate (6%) is comparable to the FEM result (4%). Indeed, the
amplitude is largely overestimated when the power is computed from the torque (>17.5%).
Figure 5.10: Evolution of the power number at Re=120 for (a) the unbaffled and (b) baffled
configurations, with
.
Figure 5.11: Variation of the power for the baffled configurations: (a) increase compared to the
unbaffled configuration and (b) maximum amplitude during one revolution.
In summary, contrary to simulations for a fixed geometry (see Section 5.2.5.3), the selection of
the method to compute the power can significantly impact the accuracy of the results around
moving solid parts. When computed from the torque, the evolution of the power in time is largely
112
perturbed by oscillations. We believe these inaccuracies are due to the initialization procedure
that is necessary when new fluid nodes appear. Conversely, when the power is computed from
the shear rate, the LBM results agree well with the experimental data and the FEM results. For
instance, the increase in power due to the baffles and the maximum amplitude that occurs during
one impeller revolution are accurately simulated.
5.2.6 Concluding remarks
The objective of this work was to assess the suitability of the extrapolation method for the LBM
simulation of viscous mixing flows involving complex geometries. To our knowledge, it is the
first time that such extended accuracy analysis is presented for such complex 3D flow problems.
The emphasis was put on the capacity to simulate accurate power consumption and pumping
capacity for both unbaffled and baffled configurations of the Maxblend mixing system.
Thanks to the simulation of the Couette flow problem, the impact of the boundary conditions at
the top and bottom was analyzed. The LBM scheme was observed to achieve second-order
accuracy for the velocity, but the norm of the shear rate and power were found to be more
sensitive to a change of boundary conditions. In fact, these inaccuracies were related to local
perturbations appearing when boundary conditions are ill-defined. These results also showed that
the power computed from the torque, contrary to that obtained from the shear rate, presents a
monotonically decreasing convergence behavior.
A highly parallel LBM scheme was next used to simulate fluid flow within the Maxblend mixing
system using a Lagrangian frame of reference. A good agreement with experimental data and
finite element simulation results was found for the power consumption and the pumping capacity.
Note that, in this case, the method used to compute the power (from the shear rate or the torque)
does not significantly impact the accuracy of the power consumption. The analysis was then
extended to determine the optimum bottom clearance for three different Reynolds numbers. The
results suggest that the standard bottom clearance is well suited for the laminar regime, but a
larger value would be more efficient for the transitional regime. Moreover, they confirm that the
LBM is an adequate tool for the optimization of design parameters.
Lastly, the results obtained for the more geometrically complex case (baffled configuration)
showed that the evolution in time of the power consumption can be affected by numerical
113
oscillations. These inaccuracies are related to the initialization procedure required for the
treatment of the new fluid nodes that appear around the moving impeller. The power computed
from the torque is more sensitive to these oscillations but, fortunately, their amplitude decreases
when the lattice spacing is reduced.
5.2.7 Acknowledgments
The computer resources and support from Calcul Québec and the financial support of the Natural
Sciences and Engineering Research Council of Canada (NSERC) are gratefully acknowledged.
We also would like to thank Pr. L. Fradette from Ecole Polytechnique de Montréal and R.
Yatomi from SHI Mechanical & Equipment Inc, for useful discussions on the Maxblend mixer.
5.3 Résultats complémentaires
Les résultats présentés dans cette section ont pour but de compléter l’étude de convergence
spatiale sur le problème du Couette cylindrique, proposée dans la section 5.2.4. Plus précisément,
nous allons nous intéresser à l’impact du facteur de relaxation
sur la précision. Dans le cas des
régimes transitoire ou turbulent, la valeur du facteur de relaxation doit être proche de 0.5 afin de
conserver un nombre de Mach en dessous de 0.2, c’est-à-dire d’assurer la stabilité numérique et
la précision du schéma LBM. Dans le cas des écoulements en régime laminaire, les nombres de
Mach correspondants étant plus faibles, il y a plus de liberté sur la sélection de la valeur de .
Nous allons nous intéresser plus particulièrement au cas de l’écoulement de Couette cylindrique
avec des conditions aux limites périodiques sur la hauteur (voir section 5.2.4.1), pour un nombre
de Reynolds de 1. Ainsi, la figure 5.12 présente les erreurs sur la vitesse et sur le taux de
cisaillement pour des facteurs de relaxation compris entre 0.6 et 1.4, les différentes courbes
représentant les erreurs associées aux quatre tailles de lattice utilisées. Il ressort des figures
5.12(a), pour la vitesse, et 5.12(b), pour le taux de cisaillement, que les tendances sont similaires
c’est-à-dire que l’erreur diminue en fonction de la taille de lattice (ordre 2, voir section 5.2.4.1) et
que la valeur de
optimale semble être située autour de 0.8 (entre 0.7 et 0.9). Avec un seul cas
test, il n’est pas possible de déterminer si la valeur optimale de 0.8 est générique ou si elle dépend
de la géométrie et des propriétés du fluide. Cependant, une valeur similaire (
) a aussi
été proposée par Zhao [147]. Basée sur une étude mathématique de LBGK et vérifiée sur
114
plusieurs cas tests, cette étude montre que cette valeur est un bon compromis entre la précision
(sur la vitesse) et la stabilité, sans être toujours optimale.
Figure 5.12 : Impact du facteur de relaxation sur la précision de (a) la vitesse et (b) du taux de
cisaillement pour le problème de Couette.
Puisque le facteur de relaxation semble affecter la précision sur la vitesse et sur le taux de
cisaillement de la même manière, il est intéressant d’étudier son impact sur d’autres grandeurs.
Ainsi, la figure 5.13 présente les erreurs sur la puissance, calculée à partir (a) du taux de
cisaillement et (b) du couple (voir section 5.2.3.2), pour des facteurs de relaxation compris entre
0.6 et 1.4. De la même manière, les différentes courbes représentent les erreurs associées aux
quatre tailles de lattice utilisées. Pour la puissance calculée à partir du couple (figure 5.13(b)), on
observe un comportement similaire à celui observé pour la vitesse et le taux de cisaillement, avec
une valeur optimale de
proche de 0.8. Dans le cas de la puissance calculée à partir du taux de
cisaillement (figure 5.13(a)), le comportement est complètement différent. En effet, comme
mentionné dans la section 5.2.4.1, l’erreur sur la puissance ne diminue pas avec la taille de lattice.
On remarque même que pour la plus petite taille de lattice (
=1.25e-3), la valeur de l’erreur
n’est plus affectée par le facteur de relaxation. Cela confirme que, avec cette façon de calculer la
puissance, l’erreur est dominée par un autre aspect.
115
Figure 5.13: Impact du facteur de relaxation sur la précision de la puissance calculée à partir (a)
du taux de cisaillement et (b) du couple pour le problème de Couette.
Comme nous l’avons mentionné dans la section 5.2.4.1, il semble que dans ce cas, le terme
dominant soit l’erreur associé à l’intégration du volume. Afin de mieux comprendre cet aspect, la
figure 5.14 présente l’erreur sur le volume en fonction de la taille de lattice. Il ressort que si le
calcul du volume est globalement convergent, avec un ordre de précision proche de 2 (1.6),
l’erreur ne diminue pas d’une façon monotone. Autrement dit, il n’est pas garanti que l’erreur sur
le volume diminue pour une petite réduction de la taille de lattice, même si le volume est
globalement convergent.
En conclusion, les résultats présentés dans cette section montrent qu’il existe une valeur du
facteur de relaxation optimale (proche de 0.8), qui permet de minimiser l’erreur sur la vitesse, le
taux de cisaillement et la puissance calculée à partir du couple. Dans le cas de la puissance
calculée à partir du taux de cisaillement, l’erreur semble être dominée par l’erreur sur le volume.
Plus précisément, comme l’erreur sur le volume ne diminue pas d’une façon monotone, la
convergence spatiale de la puissance est largement perturbée.
116
Figure 5.14: Erreur sur le volume en fonction de la taille de lattice pour le problème de Couette
cylindrique.
117
CHAPITRE 6
APPLICABILITÉ DE LA LBM POUR LES RHÉOLOGIES
COMPLEXES
6.1 Présentation du troisième article
Soumis à The Canadian Journal of Chemical Engineering le 8 Janvier 2013.
Auteurs : Vincent Stobiac, Louis Fradette, Philippe Tanguy, François Bertrand.
Cet article traite du mécanisme de pompage généré par le mélangeur Maxblend pour des liquides
newtoniens et fortement rhéofluidifiants, ces derniers étant modélisés par le modèle de CarreauYasuda (
). Plus précisément, les résultats présentés, obtenus avec la méthode de
Boltzmann sur réseau (LBM), permettent de caractériser la forme de l’écoulement dans le régime
laminaire profond. Ils permettent aussi d’apporter des éclaircissements sur des résultats
expérimentaux récents, basés sur les méthodes de décoloration et de vélocimétrie par imagerie de
particules (PIV). Tout d’abord, la variation du pompage dans le cas newtonien est étudiée et
permet de souligner les limitations de la PIV bidimensionnelle. Cette étude montre que la
variation du pompage est reliée à un changement dans la structure de l’écoulement secondaire
(composantes axiales et radiales de l’écoulement). Ensuite, l’étude sur le pompage est étendue
aux fluides fortement non newtoniens. Il ressort que le changement dans la structure de
l’écoulement secondaire, observé dans le cas newtonien, est aussi présent et qu’il entraine la
création de zones avec des fortes disparités de viscosité apparente. La forme de ces zones permet
d’expliquer les patrons d’écoulements problématiques observés expérimentalement.
118
6.2 Pumping characterization of the Maxblend impeller for Newtonian and
strongly non-Newtonian fluids
6.2.1 Summary
This paper examines the pumping mechanism generated by the Maxblend impeller for Newtonian
and strongly shear-thinning fluids, the latter based on the Carreau-Yasuda model (
).
More precisely, results obtained with the lattice Boltzmann method (LBM) are presented and
used to characterize the flow in the deep laminar regime. These LBM results revisit and
contribute to understanding recent experimental data from the decolorization method and particle
image velocimetry (PIV). First, the modification of the pumping of the Newtonian fluid in the
deep laminar regime is assessed and the limitations of the two-dimensional PIV are discussed.
This work clearly shows that the variation of the pumping is related to a change in the structure
of the secondary radial and axial flow. Second, the analysis of the pumping is extended to highly
non-Newtonian fluids and a similar transition is observed in the laminar regime. This
phenomenon leads to the creation of zones with very different apparent viscosities, thereby
explaining problematic flow patterns observed experimentally.
6.2.2 Introduction
Among the various impellers used in chemical industries, the Maxblend mixer (Sumitomo Heavy
Industries, Japan) is an interesting candidate for mixing processes involving a wide range of
viscosities. Despite a rather simple geometry, comprised of a lower large paddle and an upper
grid, good performance has been observed from the laminar to the turbulent regimes. When
compared to the helical ribbon and anchor at low Reynolds number [43, 148, 149] or to turbines
in the turbulent regime [35, 148-152], the Maxblend provides a good compromise between
mixing efficiency and power consumption. It has been successfully applied to many applications
such as polymerization [152], bacterial culture [148, 150], gas dispersion [152] and solid
suspension [151, 153].
Several experimental and numerical investigations on the Maxblend have been published
regarding the impact of design parameters in terms of power consumption [13, 33, 36, 143],
mixing time [33, 36, 143] and dispersive performance [43]. These studies have been carried out
for both Newtonian and non-Newtonian fluids. Owing to such works, the performance of the
119
Maxblend is well documented, thus confirming the advantages and drawbacks of the impeller. In
particular, if it has been established that the homogeneity of the mixture is harder to obtain at low
Reynolds number [13, 36, 143, 154], the understanding of the mixing mechanisms in the deep
laminar regime being still limited.
As the pumping capacity is a key factor for the mixing uniformity, Computational Fluid
Dynamics (CFD) has been used to characterize the performance of the Maxblend in terms of
pumping for Newtonian and non-Newtonian fluids. For instance, results obtained with the Finite
Element Method (FEM) showed that the axial flow is characterized by a downward flow around
the Maxblend and an upward flow near the tank walls [13, 36]. Although the shape of the axial
pumping volume is toroidal and rather independent of the flow regime, the pumping efficiency is
however largely affected by the Reynolds number [13, 43, 155], the rheology [13] and the bottom
clearance [13, 155]. In addition, Devals et al. [13] noticed that the pumping is reduced at low
Reynolds number and this effect seems to be amplified by the rheology. Nevertheless, they did
not give any reason for this reduction.
From an experimental standpoint, the determination of the pumping capacity is complex due to
the difficulty in measuring the flow field. However, two promising approaches, based on twodimensional Particle Image Velocimetry (PIV) [156, 157] and the so-called decolorization
method [32], have recently been utilized [35, 154]. With the 2D PIV, the flow is obtained through
a statistical analysis of successive pictures representing the positions of particles in the system.
On the other hand, the decolorization method measures the pumping indirectly. The idea is to
extract the maximum pumping capacity from a decolorization curve, while taking into
consideration that the decolorization rate is proportional to the pumping rate. The decolorization
comes from the agitation of a small quantity of a neutrally buoyant acid solution to the liquid in
the tank, which was initially rendered alkaline by the addition of a base. These recent
experimental studies on the Maxblend are highly relevant given that they investigate the mixing
of strongly non-Newtonian fluids, which are often encountered in industries (food, plastic and
cell culture) yet rarely studied by researchers. The results confirm the toroidal shape of the axial
pumping volume generated by the Maxblend impeller in the Newtonian case. They also highlight
the occurrence of problematic flow patterns, which give rise to segregation zones indicative of
inefficient mixing, for strongly shear-thinning fluids (
) at low Reynolds number.
120
The Lattice Boltzmann Method (LBM) [58, 144] has been shown to be an adequate tool for the
simulation of industrial mixing flows. In particular, this method was successfully used for the
characterization of flow fields in Rushton turbine systems [27, 86], a crystallizer with a Rushton
turbine [117], a standard anchor [140] and a Paravisc impeller [140]. Note that we recently
investigated the accuracy of the LBM in the case of the Maxblend mixing system [155]. A
thorough investigation of the simulation procedure was provided, thereby confirming the
accuracy of the LBM predictions for Newtonian fluids.
From this literature review, it appears that, in the deep laminar regime, the understanding of the
mechanisms inherent to the pumping generated by the Maxblend impeller is limited. The
objective of this work is then to shed light on these mechanisms for Newtonian and strongly nonNewtonian fluids. To do so, LBM results are first validated with previously published FEM data
[13]. These results are next used to revisit and better understand recent experimental data based
on 2D PIV [154] and the decolorization method [35]. The limitations of these experimental
approaches are highlighted and explained. To our knowledge, no such extended numerical
investigation of the pumping capacity of the Maxblend has been done for Newtonian and strongly
non-Newtonian fluids.
The paper is organized as follows: Section 6.2.3 details the geometry of the Maxblend impeller
system as well as the rheology of the fluids used in this work. The LBM is also briefly presented
along with a few comments regarding its implementation. In Section 6.2.4, results obtained with
the LBM are compared to FEM numerical results and experimental data. In the case of the
Newtonian fluid, the variation of the pumping with respect to the Reynolds number is studied by
comparing velocity profiles. Finally, in the case of the strongly non-Newtonian fluids, a similar
comparison is made and extended to the analysis of viscosity profiles.
6.2.3 Methodology
6.2.3.1 The Maxblend impeller
The mixing system used in this work is comprised of a wedged Maxblend impeller, a tank with
an elliptical bottom and two baffles, as recommended by [33]. The geometry is presented in Fig.
6.1 and its characteristics are given in Table 6.1. The impeller is composed of a large paddle at
the bottom, creating the pumping in the mixer, and an upper grid for dispersion purposes.
121
Figure 6.1: Geometry of the Maxblend mixing system.
Tableau 6.1: Geometrical characteristics of the Maxblend mixing system.
Bottom clearance
5.5 mm
D
Impeller diameter
255 mm
F
Impeller height
338 mm
H
Fluid height
410 mm
W
Baffle width
26 mm
T
Tank diameter
356 mm
V
Vessel capacity
35.4 L
6.2.3.2 Rheology of the fluids
Two different rheology models are considered in this work to simulate Newtonian and strongly
non-Newtonian behaviors. The first one models a Newtonian diluted corn syrup, which is simply
described by a constant viscosity. Following our recent experimental work [35, 154], the
122
Carreau-Yasuda model is used to describe the strongly non-Newtonian behavior of Steol
solutions containing more than 2 wt% of NaCl. The Carreau-Yasuda model [11] is defined as
[
̇
where ̇ is the shear rate (in
) defined by ̇
tensor. The parameters of the model are
(in
), respectively,
and
√
(6.1)
]
̇
⃗
, with
⃗
the shear rate
, the zero and the infinite shear rate viscosities
the characteristic time of the fluid (in ),
the power law index and
characterizing the transition between the Newtonian and shear-thinning behavior.
Tableau 6.2: Values of the rheological parameters for the Newtonian and non-Newtonian fluids.
Fluid
NaCl
concentration
(wt %)
Rheological model
Density
(
)
(Pa.s)
1
Newtonian
1390
7.2
-
-
-
-
2
Newtonian
1370
1.3
-
-
-
-
3
Carreau-Yasuda
1030
7.1
0.053
0.05
2.2
2.1
4
Carreau-Yasuda
1030
17.8
0.17
0.05
5.3
2.7
5
Carreau-Yasuda
1030
46.5
0.63
0.05
10
2.5
number
(s)
(-)
(-)
As presented in Fig. 6.2, the fluid exhibits a highly shear-thinning behavior (
exceeds the critical shear rate ̇
) when ̇
. The values of the parameters corresponding to the
different cases simulated in this work are given in Table 6.2. Throughout the rest of this paper,
the fluid number will refer to this table and to the physical properties and rheological parameters
reported therein.
123
Figure 6.2: Shear viscosity data for aqueous solutions of Steol with different salt concentrations
from 0.4 to 2.9 wt% (from [154])
In the case of the Newtonian fluids, the Reynolds number of a mixing system can be defined as
, where
),
and
are the density (in
is the rotational speed (in RPS) and
) and the viscosity of the fluid (in
is the impeller diameter (in
). For a fluid
obeying a Carreau-Yasuda model, the Reynolds number can be generalized through the concept
of Metzner and Otto [31] as
[
where
(6.2)
]
is the Metzner-Otto constant. In this work, its value is set at 15, as was proposed by
[143] for a similarly low power law index (
) and confirmed to be adequate in our
previous experimental work [35, 154].
For impeller comparison purposes, the power consumption
dimensionless characteristic numbers,
and
(in
) can be characterized by two
. The power number
, given by
(6.3)
is constant in the turbulent regime. Conversely, the power constant
laminar regime and can be easily generalized to non-Newtonian fluids as
is constant in the
.
124
The pumping capacity of a mixing system can also be represented by a dimensionless
characteristic number, the axial flow number
, which is given by
(6.4)
where the axial pumping (or flow rate)
height (in
(in
), V the total volume of the mixing system (in
), in other words,
if
∫
is computed as
and
) and
, with H the fluid
the upward axial velocity
otherwise. In the case of close-clearance
impellers, the computation of the axial pumping is generally approximated by the flow crossing a
set of horizontal planes [13, 36].
6.2.3.3 The lattice Boltzmann method
Only a brief presentation of the LBM is given here and the reader is referred to [58, 144] for
further explanations on the method and [155] for more details on the implementation used in this
work.
The lattice Boltzmann method (LBM) is a powerful numerical tool, which differs from other
classical CFD methods by obtaining the solution of the Navier-Stokes equations from a
simplified kinetic equation, the Boltzmann equation, based on the kinetic theory of gases. It has
been successfully applied to the simulation of problems highly relevant for industrial
applications, such as suspensions [72, 111, 139], turbulent flows [86, 117, 158], multiphase flows
[159-161], non-Newtonian flows [16, 27, 54, 115] and porous media flows [3, 110, 115]. The
main advantages of the LBM are its highly parallel efficiency, due to the locality of the
computations and the regularity of the mesh, as well as its ability to handle for instance
turbulence and rheologically complex fluids. The LBM simulates the macroscopic dynamics of a
fluid by propagating and colliding fluid particles along the bonds of a lattice grid. The primary
variable of the LBM is then a population of fluid particles.
Thus, at time , for a lattice at position , populations
are associated with velocity vectors
, which are determined by the type of lattice. In this work, the D3Q15 lattice [63] is used and
the velocity vectors are defined as
125
{
(6.5)
The collision-propagation procedure can then be written (with the BGK [60, 63] collision
operator) as follows
[
]
(6.6)
where the equilibrium populations are computed as
[
with
the relaxation parameter,
populations, defined as if
√
, if
(
⃗
⃗
⃗ ⃗
)]
the lattice speed of sound and
and
if
(6.7)
the weights of the
, for the D3Q15.
The macroscopic variables are computed in every lattice as
∑
(6.8)
(6.9)
⃗
where
∑
is the total number of populations per lattice and
is the mean density of the
fluid.
The viscosity of the fluid
is introduced into the method through the relaxation parameter
(6.10)
Thus, in the Newtonian case, the value of the relaxation parameter is constant in time and space,
and is used to adjust the time step
.
126
For non-Newtonian simulations, the value of the viscosity
̇ is given by a rheological model
(the Carreau-Yasuda model in this work), so that the relaxation parameter is no longer constant.
The shear rate tensor , used to compute the shear rate ̇
√
, can be evaluated from the
velocity field with a second-order centered formula. It can also be evaluated directly from the
populations [62, 89]:
∑(
where
is the dimension of the problem,
(6.11)
)
is the Kronecker delta function and
called non-equilibrium populations computed as
the so.
The latter technique is used in this work since it preserves the locality of the LBM and has shown
to provide the same accuracy for Newtonian fluids [62, 155]. In the non-Newtonian case, the
shear-rate tensor
is used to determine the relaxation parameter . It is evaluated through Eq.
(6.11) using the previous time step:
. This choice is justified by the desire to preserve
the explicit form of the LBM scheme and to prevent the use of costly fixed-point iterations.
In conclusion, the LBM is based on an explicit scheme where all computations are simple and
local. The scheme has proven to be second-order accurate in time [121, 140, 144, 145] and
recovers the solution of the Navier Stokes equations under the hypotheses of small Mach and
Knudsen numbers. With this method, these two numbers are defined as
where
|⃗ |
and
,
is a characteristic length. The impact of the limitation on the Mach number will be
discussed in the next section.
6.2.3.4 Practical aspects of the simulation process
As the LBM uses Cartesian structured meshes, the discretization of the Maxblend impeller
implies determining which lattice nodes are fluid and which ones are solid. To do so, a mesh is
used to describe the geometry of this solid part. In addition, due to its rotation, the state of the
lattices (solid or fluid) evolves with time. As the initialization of the new fluid lattices is nontrivial and may affect the accuracy of the results, we chose here to use the non-equilibrium
corrected refill procedure [88] to accomplish this task. With this approach, the missing
127
populations are determined by extrapolation, thereby reducing errors on the flow field around the
moving impeller. The reader is referred to [155] for detailed explanations. In this work, the lattice
spacing
was set to 1.77 mm, corresponding to a total close to 7 million lattices. This lattice
spacing was observed to be fine enough to yield preliminary results with variations on the
average velocity that are lower than 1%, when compared to the solutions obtained with finer
lattices.
After the discretization process, boundary conditions have to be imposed between fluid and solid
lattice nodes. Here, due to the complexity of the geometry, boundary conditions are a key aspect
of the whole simulation process. The boundary conditions for the tank walls, the baffles and the
Maxblend impeller are imposed with the Extrapolation method (EM) [78]. Owing to this
approach, the accuracy of the LBM is conserved [78, 155] and the shape of the boundaries is
precisely described. On the other hand, the free surface is approximated by a free slip boundary
condition [58], and considered to be flat. This approximation is generally used to simplify the
simulation process, as it prevents the need to simulate a two-phase flow (air and liquid). Here, the
error generated by this simplification is insignificant considering that baffles largely reduce the
height of the vortex, which in any case is small in the laminar regime.
For a given set of physical parameters (density , viscosity
⃗ ) and a given lattice spacing
and rotational speed
or velocity
, the last step of the simulation process is the selection of the
relaxation parameter . Substituting the lattice speed of sound into Eq. (6.10) leads to:
(6.12)
and the Mach number becomes:
√ |⃗ |
It appears from these two equations that the time step
Mach number are proportional. Given that the value of
(6.13)
, the relaxation parameter
is constant in the Newtonian case, its
selection is a compromise between accuracy (the smaller
(not too small
and the
). In the case of high velocity | ⃗ | (high
the better) and computational time
),
must be small enough to
guarantee a Mach number lower than 0.2, which has been observed for various problems to be
128
adequate in terms of numerical stability. This can be achieved by selecting a relaxation parameter
close to 0.5, thereby drastically reducing the range of acceptable values of .
In the case of a non-Newtonian fluid, the value of the relaxation parameter
̇ . In the case of a shear-thinning fluid (
apparent viscosity
is a function of the
), the value of this viscosity is
inversely proportional to the shear rate ̇ , as deduced from Eq. (6.1). Moreover, Eq. (6.10) shows
that the relaxation parameter
and
̇ are proportional, so that
will also decrease with an
increase in the shear rate. In this case, the selection of the reference relaxation parameter
(
̇
) is crucial, as it must be high enough to account for the total variation
of the viscosity in the computational domain during the simulation. In this work, as the fluid is
strongly shear-thinning (
),
was fixed at 4. In that the relaxation parameter
Mach number are proportional (see Eqs. (6.12) and (6.13)), for a given value of
and the
(1.77 mm in
this work), the velocity ⃗ must be small enough to ensure a small Mach number. Basically, the
more the fluid is shear-thinning, the smaller the maximum achievable Reynolds number. Note
that, for a given Mach number, the impact of the rheology on the computational time is
negligible. In addition, preliminary tests in the case of the circular Couette flow with a power-law
model have shown that the second-order accuracy of the LBM was conserved.
Our parallel implementation of the LBM is based on the use of the MPI library. All the
computations were performed on Calcul Quebec Colosse supercomputer, the characteristics of
which are given in Table 6.3.
Tableau 6.3: Characteristics of the Colosse supercomputer.
Number of nodes
960
Type of node
2 Intel Nehalem-Ep 4-core proc.
Total number of cores
7680
Memory per node
24 GB
Total available memory
23 TB
Network
Infiniband QDR (40 Gb/s)
Peak performance
88 Tflops
129
6.2.4 Simulation results
In this section, LBM simulations are presented for Newtonian and strongly shear-thinning fluids
(Carreau-Yasuda model with
). In order to validate the simulation process, results are
first compared to experimental data and FEM results.
6.2.4.1 Newtonian flow
First, we focus on Newtonian results in order to validate the simulation process, as more results
are available in the literature for this type of fluids. Our LBM results are compared with FEM
results [13] and data from 2D PIV measurements [154] and the decolorization technique [35]. For
the FEM simulations, an unstructured mesh of 290,336
tetrahedral elements was used.
The corresponding system of 2.12 million equations was solved by means of the Krylov-based
Uzawa solver.
The LBM results are presented for Reynolds numbers up to 140, as our current LBM simulation
procedure became numerically unstable for higher
with
mm. This limitation could
be overcome by reducing the lattice spacing or using the MRT collision model [122] to improve
the numerical stability. Nevertheless, going above
was not deemed necessary within
the scope of this work. The corresponding maximum Mach numbers were lower than
for
all LBM simulation results, which is small enough to ensure a good accuracy. In addition, the
total number of impeller revolutions required to achieve a periodic solution depends on the
Reynolds number. For instance, at
revolutions were required for
, only 2 revolutions were necessary. Conversely, 12
. The average computational time for 1 revolution was
around 3.5h with 48 cores.
6.2.4.1.1 Power and pumping curves
The
versus
power curves, obtained from experimental data based on torque measurements
[35], FEM results [13] and our LBM simulations are presented in Fig. 6.3(a). A good agreement
can be noticed between the LBM and FEM results and the experimental data. First, in the laminar
regime (
), the slope is constant at -1, confirming that the power number is linearly
proportional to the Reynolds number (i.e.
). Next, for higher Reynolds
numbers, a transition zone can be observed as the slope of the curve decreases.
130
Figure 6.3: Power and pumping curves for the Maxblend mixing system in the Newtonian case
with experimental data [35], FEM [13] and LBM results.
In the case of the pumping capacity, experimental measurements are generally more difficult to
take, although recent approaches appear to be promising. For instance, the use of the
decolorization method [32] has recently been extended to evaluate flow number
from the
decolorization rate [35]. From this information, an estimate of the global pumping (radial,
tangential and axial) is obtained. The value of the axial pumping is therefore overestimated, but
the trend with respect to the Reynolds number has shown to comply with that obtained using
other techniques.
In order to compare the axial pumping predicted by the LBM to the global pumping obtained
from the decolorization method, it is then necessary to scale the results. To do so, a normalized
flow number is introduced:
(6.14)
where
and
are the minimum and maximum flow numbers for the corresponding
techniques.
Fig. 6.3(b) displays the variation of the normalized flow number
versus the Reynolds number
from the decolorization method [35] as well as the FEM [13] and LBM results. A good
agreement can be noticed between the two numerical approaches as the difference is generally
131
under 5%, except at Reynolds numbers close to 1 where the differences are closer to 10%.
Comparisons with the experimental data are also good as the three curves exhibit the same trend:
a small pumping capacity in the deep laminar regime (
the transition regime up to
), followed by a sharp increase in
. However, the decolorization method is not accurate
enough to describe the pumping efficiency in the deep laminar regime, let alone the slight
increase predicted by both the LBM and FEM when
. The latter numerical results
suggest that the axial flow may be different in this regime. The next section will investigate this
aspect.
6.2.4.1.2 Analysis of the flow field
The flow generated by an impeller can be decomposed into two parts: the tangential component
that follows the rotation of the impeller and the secondary flow, which is the sum of the axial (i.e.
pumping) and radial contributions. As the distributive mixing mechanism is principally generated
by the secondary flow, it has already been investigated experimentally and numerically in
previous investigations focusing on the Maxblend [13, 154]. Recently, the 2D PIV was applied to
investigate the effect of highly shear-thinning and viscoelastic fluids on the secondary flow [154].
These results revealed the dependence of the secondary flow patterns on the fluid rheology.
However, one has to keep in mind that those measured values of the radial and axial components
of the flow come from the projection of the flow field on a plane perpendicular to that goes
through the impeller and the baffles. While it is known that perspective errors in 2D PIV can be
significant [162], the data reported in [154] seemed to capture qualitatively well the flow fields at
various Reynolds numbers.
Fig. 6.4 compares the secondary flows obtained with the results of the LBM, PIV [154] and FEM
[13], for Reynolds numbers of 2, 23, 55 and 140. The vertical cross-sections are perpendicular to
the impeller and represent only one half of the domain. It is important to mention that the
magnitude of the secondary flow is different for the FEM results because they pertained to other
viscosities and densities. In addition, the studied mixing system was comprised of 4 baffles rather
than 2.
132
Figure 6.4: Magnitudes (Vrz) of the radial and axial components of the velocity for the
Newtonian fluid obtained with the LBM, PIV [154] and FEM [13].
133
For Reynolds numbers between 23 and 140, Figs 6.4(b), (c) and (d) show a rather good
agreement between the secondary flows obtained with the three approaches. These results are in
accordance with previous works [13, 35, 36, 154], in that the shape of the secondary flow does
not drastically change over this range of
. More precisely, it is characterized by an intense
downward flow around the impeller and a moderate upward flow near the tank walls, which is
reminiscent of a toroidal shape. In addition, when comparing the 2D PIV and LBM results for the
same configurations (density, viscosity and impeller speed), it appears that the magnitude of the
maximum velocity obtained with both approaches are similar. In order to extend this analysis,
Fig. 6.5 presents the profile of the dimensionless axial velocity
radius, at
with respect to the
for different Reynolds numbers from 6 to 55. Once again, a good
agreement can be observed among the results obtained with the LBM, PIV [154] and FEM [13].
In particular, it confirms that the maximum axial velocities measured with the 2D PIV are close
to the numerical results.
Figure 6.5: Dimensionless axial velocity profiles at z/H=0.75 with the LBM, PIV [154] and FEM
[13].
In the case of the secondary flows obtained at
, Fig. 6.4(a) shows differences in the
contours of the secondary flows obtained with the LBM, the FEM and the PIV. For instance, a
good agreement can be noticed between the numerical results, although the upper part of the
secondary flow is slightly underestimated by the LBM. We consider that the most likely
134
explanation for this lies in the difference in the Reynolds number (6 for the FEM and 2 for the
LBM) and the differences in the mixing systems as mentioned before. Larger discrepancies can
be noted with the 2D PIV contours. While similarities do exist between the PIV and LBM results
regarding the shape of the flow, the magnitude of the maximum velocity is 9 times larger with the
PIV. This is a priori surprising given that for other Reynolds numbers, the maximum velocities
were similar for these two methods. We will return to this point in the next section.
From a broader perspective, the secondary flows obtained at
with the LBM, the PIV and
the FEM, suggest that the toroidal shape of the axial pumping volume observed at higher
is
not valid in the deep laminar regime. This result is compliant with conclusions drawn in the
previous section, in that the pumping was not constant in the laminar regime (see Fig. 6.3(b)).
Clearly, a better characterization of the secondary flow in the deep laminar regime can help
understand the variation of the pumping efficiency found therein and the related mixing issues
raised in previous works [13, 36, 154].
6.2.4.1.3 Transition from the deep laminar to the transitional regimes
In previous numerical and experimental studies [13, 154], the characterization of the secondary
flow was limited to the visualization of sections perpendicular to the impeller. This choice can be
easily explained by the difficulty to experimentally measure the flow in other positions, as the
impeller then hides the view. However, we will show that resorting to only these cross-sections in
the deep laminar regime may lead to inaccurate perceptions.
Fig. 6.6 displays the three-dimensional shape of the axial pumping volume obtained with the
LBM, for Reynolds numbers of 2 and 23. In both cases, the negative values of the axial
component of the velocity are in blue and the positive ones in red. Figs. 6.6(c) and 6.6(d) show
the standard toroidal shape of the axial pumping volume generated by the Maxblend at
.
As mentioned previously, the corresponding secondary axial and radial flow is composed of a
downward component around the impeller and an upward component near the tank walls.
Conversely, Figs. 6.6(a) and 6.6(b) show that the axial flow does not follow the same behavior at
. In fact, the fluid goes up in front of the rotating impeller and goes down behind. We
believe this phenomenon is due to the predominance of the viscous forces, which prevent the
formation of the shape observed at the higher value of
. This is worthwhile mentioning since it
is the first time that the secondary flow is shown to be completely different in the deep laminar
135
regime. It also explains why poor mixing performances have been reported in this regime [13, 36,
143, 154].
Figure 6.6: Axial pumping volumes of the Newtonian flow predicted with the LBM, at Reynolds
numbers of 2 and 23: (a) and (c) correspond to three-dimensional views and (b) and (d) to top
views. The red and blue colors denote positive and negative values of the axial component of the
velocity, respectively. The dashed lines represent the section used for the PIV measurements, and
the arrows indicate the sense of rotation.
Figs. 6.6(b) and (d) show top views of the axial pumping volumes at Reynolds numbers of 2 and
23, respectively. The dashed lines represent the section used for the 2D PIV measurements and
the secondary flow fields in Fig. 6.4. Fig. 6.6(b) clearly shows that, at
, using such a
section is inadequate to characterize the secondary flow in the tank, as the volumes are not
axisymmetric and the absolute values of the velocities are very small. It explains why the 2D PIV
fails to measure a precise maximum velocity in this case. Conversely, for higher Reynolds
numbers, when the shape of the axial pumping volume is toroidal, the section perpendicular to
136
the impeller is adequate to characterize the secondary flow. These results illustrate that
simulation can help set up the sections used for 2D PIV measurements. It also underlines the
limitations of 2D flow visualization.
Knowing that the secondary flow is different in the laminar regime, it is then important to
determine how it affects the pumping efficiency. Fig. 6.7 presents the variation of the relative
importance of the three components of the time-averaged velocity. It appears from the curves that
the radial component of the flow is rather small and constant (from 9.5 to 12%). Next, for the
axial component, the curve can be decomposed into three parts: a rather small value (lower than
13%) in the deep laminar regime (
), followed by a linear increase in the laminar regime,
and finally, a rather constant but large value (around 24%) for
. These results establish
that the conversion of the tangential motion into pumping (or axial motion) is weak and
minimum around
. This minimum corresponds to a state in between the two secondary
flows depicted in Figs. 6.6(b) and (d). For higher Reynolds numbers (i.e.
), the
conversion of the tangential motion into pumping increases, thereby indicating that in the case of
a toroidal secondary flow (Figs. 6.6(c) and (d)), the increase of the inertial forces leads to
enhanced pumping. The relative intensity of the axial flow seems to level off at around
Figure 6.7: Variation of the relative importance of three components of the time-average mean
velocity obtained with the LBM in the Newtonian case.
.
137
6.2.4.2 Strongly non-Newtonian flow
In this section, we extend our investigation to a strongly shear-thinning fluid, in order to evaluate
the impact of the rheology on the pumping capacity of the Maxblend impeller. The maximum
Mach numbers achieved were lower than 0.15, which is sufficiently small to ensure a reasonable
accuracy. As the fluid is strongly shear-thinning (
), a minimum viscosity of
was set in order to prevent the occurrence of instabilities. Such strongly non-Newtonian
fluids can be easily tackled with the LBM, but the reference relaxation parameter
must be
cautiously chosen (see Section 6.2.3.4). Note that our FEM code was unable to simulate a power
low index
lower than 0.3 [13].
Once again, the total number of impeller revolutions required to achieve a periodic solution
depends on the generalized Reynolds number. Given that the rotational speeds are rather small, a
maximum of 5 revolutions was necessary at a generalized Reynolds number
of 55. The
average computational time for 1 revolution was around 4h on 48 cores.
6.2.4.2.1 Power and pumping curves
Fig. 6.8(a) displays the power curves obtained with non-Newtonian LBM simulations (fluid no. 5
in Table 6.2) and experimental data [35] for Newtonian and non-Newtonian fluids. As suggested
in [35], the use of the generalized Reynolds number
, coupled with a Metzner-Otto constant
of 15, allows to use the Newtonian power curve as a master curve. Furthermore, even if the fluid
is strongly shear-thinning (
), a good agreement is observed between LBM results and
experimental data.
In addition, Fig. 6.8(b) shows the variation of the axial flow number
with respect to the
Reynolds number (using fluids nos. 1, 2 and 5 presented in Table 6.2). Quite clearly, the pumping
capacity of the Maxblend is largely affected by the rheology of the fluid. It is important to
mention that this result is limited to the laminar and transitional regimes, as the pumping was
recently shown to be independent of the rheology in the turbulent regime [35]. In the latter work,
the occurrence of problematic flow patterns giving rise to two segregated regions in the tank were
revealed by the decolorization method. As an illustration, Fig. 6.9 shows these two segregated
zones for
. To gain insight into the effect of rheology on flow behavior in the laminar
regime, the flow field of the strongly shear-thinning fluid is next analyzed.
138
Figure 6.8: Power and pumping curves for the Maxblend mixing system: experimental data [35]
and LBM results for the (N) Newtonian and (NN) non-Newtonian cases.
Figure 6.9: Segregation zones for
. Adapted from [35].
6.2.4.2.2 Analysis of the flow field
Fig. 6.10 compares the contours of the secondary flows obtained with the 2D PIV [154] and
LBM, for generalized Reynolds numbers of 6 and 9, corresponding to fluids nos. 3 and 4,
respectively (see Table 6.2). While the lower parts are rather similar in all cases, the upper parts
of the contour plots exhibit different behaviors. In addition, the 2D PIV overestimates the
139
magnitude of the secondary flow (radial and axial components of the velocity), but the
differences decrease with the increase in
(from a factor of 10 at
to 4 at
is noteworthy that the agreement between the LBM and PIV are not as good at low
). It
, as in this
case, the absolute values of the measured velocities are too small to ensure a good accuracy, as
discussed in Section 6.2.4.1.3. Similarly to the Newtonian results (see Fig. 6.6(b)), the position of
the section used for the 2D PIV measurements is inadequate.
Figure 6.10: Magnitudes (Vrz) of the radial and axial components of the velocity for the strongly
non-Newtonian fluid obtained with the LBM and PIV [154] for (a)
(fluid 3) and (b)
(fluid 4).
In summary, the LBM can be successfully used to characterize the secondary flow with strongly
non-Newtonian fluids. However, with the current BGK collision model and a lattice spacing
mm, the maximum
that can be simulated is limited, as mentioned in Section
6.2.3.4. The 2D PIV remains a good alternative in the transitional and turbulent regimes, as it was
shown to accurately determine the shape of the secondary flow as well as the maximum value of
the axial velocity.
6.2.4.2.3 Transition from the deep laminar to the transitional regimes
140
Figure 6.11: Axial pumping volumes of the strongly non-Newtonian flow predicted with the
LBM at
of (a) 1, (b) 15, (c) 30 and (d) 50. The red and the blue colors denote positive and
negative values of the axial component of the velocity, respectively.
141
Simulations up to
were achieved using fluid no. 5 (see Table. 6.2) to investigate the
variation of the pumping efficiency with respect to
. Three-dimensional shapes of the axial
pumping volume obtained with the LBM are presented in Fig. 6.11, which correspond to
generalized Reynolds numbers of 1, 15, 30 and 50, respectively. In both cases, the negative
values of the axial component of the velocity are in blue and the positive ones in red. Fig. 6.11(a)
shows that the axial flow at
is rather similar to the one observed at
, in the
Newtonian case. This flow is characterized by an upward component in front of the rotating
impeller and a downward component behind. The similarity with the Newtonian case implies
that, in the entire domain, the local shear rate is lower than the critical shear rate ( ̇
the fluid. With the increase of
) of
, it can be observed in Figs. 6.11(b-d) that the shape of the
axial pumping volume is significantly modified. Similarly to the results for the Newtonian fluid,
the fluid flows downward around the impeller and upward near the tank walls. However, contrary
to the Newtonian case, the pumping in the upper part of the mixer is much weaker compared to
the one at the bottom. It shows that, in case of a strongly non-Newtonian fluid, even if the paddle
(located at the bottom) is able to generate pumping, the grid (located at the top) fails to conserve
it. In order to determine if this lack of pumping around the grid can be related to the problematic
flow patterns observed experimentally at
[35] (see Fig. 6.9), the variation of the
apparent viscosity in the tank is analyzed.
Fig. 6.12 gives the maps of the viscosity obtained with the LBM, for Reynolds numbers of 1, 15,
30 and 50. The vertical cross sections are perpendicular to the impeller and the baffles. Fig.
6.12(a) confirms that, at
, the shear rate is globally under the critical value of the fluid, as
the viscosity is hardly affected by the motion of the impeller. In fact, the bottom of the tank is the
only area where the viscosity decreases, as the shear rate is maximal in the small clearance
between the impeller and the tank. For higher generalized Reynolds numbers, as shown in Fig.
6.12(b-d), an envelope of small viscosities develops, above which the viscosity becomes rapidly
large and within which a number of isolated large viscosity islands can be noticed. These results
are relevant as the shape of the top of the envelope is similar to the interface between the
segregated zones observed in Fig. 6.9. In addition, as the viscosity at the top of the tank is high
for every Reynolds number, it shows that even if the paddle does impart rather large shear rates at
the bottom of the tank, the upper part of this paddle and the grid are unable to do so. As a matter
142
of fact, this phenomenon is reduced with an increase of the generalized Reynolds number, as the
size of the islands diminishes and the envelope occupies more and more space in the tank.
Figure 6.12: Viscosity maps obtained with the LBM for the strongly non-Newtonian fluid at
of 1, 15, 30 and 50.
As a complement, Fig. 6.13(a) presents the variation of the relative importance of the three
components of the time-average mean velocity with respect to the generalized Reynolds number.
It shows that the radial and axial components are both small (always lower than 13%) for all
.
It suggests that, in this range of generalized Reynolds numbers, increasing the inertia forces does
not create more pumping. Fig. 6.13(b) compares the axial components of the time-average mean
velocity for the Newtonian and highly shear-thinning fluids (from Figs. 6.7 and 6.13(a)). It
appears that the pumping of the strongly non-Newtonian fluid is weaker. Moreover, the Reynolds
number corresponding to the minimum value of the axial component is slightly larger (20 as
opposed to 10), which is close to 15, the value at which the volume of the islands was maximum
in Fig. 6.12. For complex mixing applications, it is important to determine efficient operating
conditions. The curves in Fig. 6.13 imply that, in the case of a strongly non-Newtonian fluid, a
143
range of Reynolds numbers (from
to 30) larger than that for a Newtonian fluid (from
to 15) should be avoided to prevent mixing problems.
Figure 6.13: (a) Variation of the relative importance of the three components of the time-average
mean flow for the non-Newtonian case and (b) comparison of the axial components for the
Newtonian and non-Newtonian fluids.
6.2.5 Conclusions
The objective of the work was to shed light on the pumping generated by the Maxblend impeller
for Newtonian and strongly non-Newtonian fluids. To do so, LBM results were first compared to
previous FEM simulations [13] in the Newtonian case. These results were also used to explain
recent experimental data based on 2D PIV [154] and the decolorization method [35]. In
particular, the importance of the section used for the 2D PIV measurements was underlined. In
the case of a strongly non-Newtonian fluid (
), the LBM was shown to be efficient in so
far as it was able to simulate the flow of this type of fluid contrary to our finite element code.
First, in the Newtonian case, we showed that the LBM results are in good agreement with the
FEM results and the experimental data. From this comparison, it was noticed that the pumping
capacity of the Maxblend impeller is highly affected by the Reynolds number. Thanks to threedimensional visualizations, the variation of the pumping efficiency in the laminar regime was
related to a change of structure in the secondary flow. In fact, in the deep laminar regime, the
axial flow is characterized by an upward component in front of the rotating impeller and a
144
downward component behind. This suggests that, under a critical Reynolds value (
), the
inertia is too small to transform the tangential flow into pumping. In the case of strongly nonNewtonian fluid flows, the same type of transition from radial to axial flow was observed in the
laminar regime. However, contrary to the Newtonian flow, a lack of pumping was noticed in the
upper part of the impeller. Finally, the presence of an envelope of small viscosities was
underlined and related to the creation of two segregated zones that have been observed
experimentally [35].
6.2.6 Acknowledgments
The computer resources and support from Calcul Québec as well as the financial support of the
Natural Sciences and Engineering Research Council of Canada (NSERC) are gratefully
acknowledged. We also would like to thank A. Fontaine and Y. Gunztburger for fruitful
discussions on the Maxblend mixer and experimental methods.
6.3 Résultats complémentaires
Les résultats présentés dans cette section ont pour but de vérifier les propriétés de convergence
spatiale du schéma LBM dans le cas de fluides non newtoniens. Pour ce faire, nous allons étudier
à nouveau l’écoulement de Couette cylindrique. Dans cette section, les conditions aux limites sur
la hauteur sont périodiques (voir section 5.2.4.1) et le nombre de Reynolds vaut 1. La rhéologie
̇
du liquide est modélisée par la loi de puissance
, avec
. Dans ce cas, les
solutions analytiques en vitesse, en taux de cisaillement et en puissance sont :
(
‖ ̇‖
)
(
)
(6.16)
√
(
(
(6.15)
)
)
(6.17)
(
)
145
La figure 6.14 présente ainsi les erreurs sur (a) la vitesse et (b) le taux de cisaillement en fonction
de la taille de lattice, pour des indices de puissance de 0.2, 0.4, 1 (newtonien) et 1.4. Il ressort de
la figure 6.14(a) que si la valeur de l’indice de puissance n’affecte pas l’ordre de précision en
vitesse, les erreurs augmentent lorsque
diminue. La conservation de l’ordre 2 de précision n’est
pas surprenante puisque le modèle rhéologique n’affecte pas le développement mathématique du
passage de Boltzmann à Navier-Stokes [14], présenté dans la section 2.1. La conservation de
l’ordre 2 en vitesse a d’ailleurs déjà été soulignée par Boyd [15] pour un modèle de loi de
puissance. De la même manière, la figure 6.14(b) montre que la valeur de l’indice de puissance
n’affecte pas l’ordre 2 de précision du taux de cisaillement et que les erreurs augmentent lorsque
diminue. On notera quand même que la stabilité numérique est plus sensible à la rhéologie
puisque, pour la taille de lattice la plus petite, les simulations ne convergent pas pour
Figure 6.14: Impact de l’indice de puissance
.
sur la précision de (a) la vitesse et (b) du taux de
cisaillement pour le problème de Couette.
Il est aussi important de vérifier l’impact de la rhéologie sur la précision de la prédiction de
consommation de puissance. Ainsi, la figure 6.15 présente les erreurs sur la puissance, calculées à
partir (a) du taux de cisaillement et (b) du couple (voir Section 5.2.3.2), pour des indices de
puissance de 0.2, 0.4, 1 (newtonien) et 1.4. Pour la puissance calculée à partir du couple (figure
6.15(b)), on observe un comportement similaire à celui observé pour la vitesse et le taux de
cisaillement, c’est-à-dire un ordre de précision proche de 2 et une augmentation de l’erreur
lorsque
diminue. Dans le cas de la puissance calculée à partir du taux de cisaillement (figure
146
6.15(a)), comme mentionné dans la Section 5.2.4.1, l’erreur sur la puissance ne diminue pas
forcément avec la taille de lattice.
Figure 6.15: Impact de l’indice de puissance
sur la puissance calculée à partir (a) du taux de
cisaillement et (b) du couple pour le problème de Couette.
147
CHAPITRE 7
DISCUSSION GÉNÉRALE
Les écoulements visqueux sont caractérisés par une forte résistance à l’écoulement, ce qui
empêche la création de tous phénomènes turbulents. Un des exemples les plus classiques
concerne les écoulements dans les milieux poreux, caractérisés par une surface de contact
importante et de géométrie complexe. Dans le cas général, la forme de l’écoulement n’est pas
seulement liée à la géométrie du problème, elle est aussi fortement affectée par la rhéologie du
fluide. Ces problématiques sont encore plus importantes dans le cas des écoulements
polyphasiques. En effet, à la géométrie du problème s’ajoutent celles des frontières qui séparent
les phases. La forme de l’écoulement n’est plus seulement affectée par les propriétés des
différentes phases, mais aussi par leurs interactions. Dans cette thèse, nous nous sommes
intéressés plus particulièrement au cas des écoulements monophasiques visqueux en cuves
agitées.
Les mélangeurs spécialisés pour les liquides visqueux sont caractérisés par des géométries
complexes, nécessaires à l’homogénéisation du mélange en l’absence de turbulence. Dans le cas
des mélangeurs discontinus, les pales doivent racler la cuve afin de prévenir la stagnation du
produit près des parois. La performance d’un système de mélange est alors associée à sa capacité
à disperser et à distribuer, ces grandeurs étant quantifiées à partir de la consommation de
puissance et de la création de pompage. Si les performances sont largement dépendantes de la
géométrie des pales et des cuves, la rhéologie de la mixture joue aussi un rôle prépondérant.
En complément des méthodes expérimentales, l’utilisation des méthodes numériques permet
d’améliorer la compréhension des écoulements visqueux en cuves agitées. Afin de prédire
adéquatement l’hydrodynamique au sein d’un système de mélange spécialisé pour les liquides
visqueux, l’interaction entre le liquide et la pale doit être précisément modélisée. Autrement dit,
la méthode numérique doit permettre un traitement précis de la géométrie, des parties mobiles et
de la rhéologie du liquide. La littérature montre que c’est principalement la méthode des éléments
finis qui est utilisée dans ce cas. La problématique de la complexité de la géométrie est
généralement traitée par une méthode de conditions aux limites immergées. Comme les modèles
rhéologiques viennent perturber le système d’équations obtenu, la majorité des résultats concerne
des écoulements monophasiques dont les propriétés rhéologiques ne sont pas fortement non
newtoniennes (
). En conclusion, les choix de modélisation sont généralement basés sur
148
un compromis entre la complexité et les temps de calcul, ce qui limite le passage à l’étude
d’écoulements polyphasiques ou à rhéologies complexes. Il est alors pertinent de tester d’autres
méthodes numériques afin de dépasser ces limitations. Ainsi, l’objectif de cette thèse était de
développer une méthode de Boltzmann sur réseau efficace pour la simulation des écoulements
visqueux en cuves agitées.
La méthode de Boltzmann sur réseau (LBM) est une approche récente, basée sur la théorie
cinétique des gaz. L’utilisation de l’échelle mésoscopique permet de modéliser la dynamique
d’un fluide à partir du comportement d’un groupe de particules imaginaires. De cette manière, la
solution des équations de Navier-Stokes est obtenue par l’application d’un schéma explicite, basé
sur des calculs simples et locaux. Associé à un maillage structuré cartésien, la méthode est
hautement parallèlisable et permet l’utilisation d’un très grand nombre d’éléments ou de cellules
(lattices). La vision mésoscopique du fluide facilite l’intégration de phénomènes microscopiques
comme la rhéologie, mais présente aussi certains désavantages. En effet, l’utilisation d’une
variable primaire mésoscopique complique l’imposition de grandeurs macroscopiques, ce qui
rend difficile l’imposition de conditions aux limites et l’initialisation des lattices.
Les différentes applications de la LBM dans la littérature permettent de tirer plusieurs
conclusions. Tout d’abord, les simulations d’écoulements dans les milieux poreux montrent que
la LBM permet de traiter un niveau de complexité géométrique difficilement atteignable avec les
méthodes classiques de la CFD. Ensuite, pour des écoulements simples avec des solutions
analytiques, plusieurs études montrent la sensibilité de la précision à la définition des conditions
aux limites. De la même manière, l’initialisation des lattices peut affecter la précision, notamment
autour des géométries avec parties mobiles. Finalement, les résultats pour des systèmes de
mélange comprenant des turbines Rushton prouvent que la LBM est efficace pour la simulation
du mélange en régime turbulent. Ils montrent aussi que l’intégration de modèles rhéologiques et
d’approches Euler-Lagrange, comme les méthodes de particules discrètes (DPM), est réalisable
dans des petits systèmes de mélange. Ces conclusions ont permis de définir trois objectifs
spécifiques reliés à l’applicabilité de la LBM aux écoulements visqueux en cuves agitées.
Le premier objectif spécifique était de comparer les différentes stratégies pour l’imposition des
conditions aux limites en LBM. Cette étape est très importante puisque le mélange visqueux est
généré par l’interaction entre la pale et la cuve, ce qui nécessite une modélisation précise de leurs
149
effets sur le fluide. Cependant, l’utilisation d’une variable primaire mésoscopique complique
l’imposition de grandeurs macroscopiques comme la vitesse de l’obstacle. Il est donc très
important de comparer les précisions qui peuvent être obtenues avec les stratégies proposées dans
la littérature. Pour ce faire, trois stratégies ont été sélectionnées parmi les approches applicables
aux géométries complexes : La méthode du rebond modifié (MBB), la méthode d’extrapolation
(EM) et la méthode de conditions aux limites immergées pour la LBM (IB-LBM). Chacune de
ces stratégies est représentative d’une classe de conditions aux limites possédant les mêmes
caractéristiques. Ainsi, la stratégie MBB est la plus simple car elle approxime les frontières
courbes par des profils en escalier. La stratégie EM utilise la distance à la paroi pour extrapoler
les populations aux nœuds solides. Finalement, la stratégie IB-LBM utilise des points de contrôle
pour contraindre la vitesse des nœuds environnants.
Dans l’article 1, les comparaisons ont été effectuées sur des géométries fixes sans partie mobile,
afin d’éviter les erreurs associées à l’initialisation des nœuds nouvellement fluides qui
apparaitraient atour d’un obstacle en mouvement. Tout d’abord, la simulation de l’écoulement de
Couette cylindrique a permis de vérifier que seule la stratégie EM est capable de conserver la
précision du deuxième ordre pour la convergence spatiale de la vitesse. Dans le cas des stratégies
MBB et IB-LBM, la précision chute au premier ordre. Cette perte de précision globale est due à
la création d’erreurs importantes près des frontières, plus particulièrement à celles où sont
imposées des vitesses non nulles. La simulation de l’écoulement de Couette cylindrique a aussi
permis d’étudier la sensibilité de la précision aux paramètres numériques des stratégies EM et IBLBM. Dans le cas de la stratégie EM, il ressort que l’impact des coefficients d’extrapolation est
négligeable et que leur optimisation n’est pas essentielle. De la même manière, l’introduction
d’une erreur sur le calcul des distances n’affecte pas significativement la précision globale. À
l’inverse, la précision des résultats obtenus avec la stratégie IB-LBM est très sensible à la
répartition des points de contrôle. Il ressort que la variabilité des résultats peut être importante
pour une petite variation du nombre ou de la position des points de contrôle. Finalement, il a été
montré que les stratégies d’imposition de conditions aux limites affectent aussi la stabilité
numérique. Plus précisément, la stratégie IB-LBM ne permet pas d’atteindre des nombre de Mach
supérieurs à 0.1, tandis que les stratégies MBB et EM sont numériquement stables jusqu’à 0.3.
La comparaison des stratégies d’imposition de conditions aux limites a ensuite été étendue à des
systèmes de mélange comprenant une ancre et un Paravisc. Afin d’éviter les erreurs associées à
150
l’initialisation des nœuds nouvellement fluides, les simulations ont été effectuées dans un repère
Lagrangien fixé sur la pale. Les stratégies ont été comparées sur les prédictions de consommation
de puissance et de capacité de pompage. Les résultats sur la consommation de puissance montrent
que les stratégies MBB et EM donnent des résultats très similaires, proches des valeurs
disponibles dans la littérature. De plus, la diminution de la taille des lattices permet d’observer
une convergence vers une valeur fixe, proche des données expérimentales. Notons que, les
résultats obtenus avec la stratégie IB-LBM sont différents de ceux obtenus avec MBB et EM,
mais restent acceptables. Les résultats sur la capacité de pompage permettent de compléter la
comparaison puisqu’ils confirment une proximité des prédictions obtenues avec les stratégies
MBB et EM. À l’inverse, la stratégie IB-LBM donne des résultats inférieurs, avec une variabilité
conséquente en fonction de la taille de lattice. En conclusion, dans le cas des mélangeurs raclants,
il semble que l’utilisation de la stratégie IB-LBM ne permet pas d’obtenir le même degré de
précision qu’avec les deux autres approches. L’observation des champs de vitesse montre que si
l’écoulement global est bien représenté, des perturbations apparaissent aux alentours de la pale,
ce qui peut générer des erreurs locales importantes sur la vitesse. Il est aussi important de noter
que la proximité des résultats obtenus avec les stratégies MBB et EM est plutôt surprenante,
puisqu’elles présentaient des précisions différentes sur l’écoulement de Couette cylindrique. Pour
autant, la comparaison reste partielle puisqu’elle ne concerne que des grandeurs globales. Il est
donc possible que l’étude d’autres aspects souligne des différences notables entre les deux
stratégies. Dans le cas des mélangeurs industriels, il serait intéressant d’étudier l’impact des
stratégies sur des configurations présentant des zones mortes, afin d’étudier les prédictions sur
leurs positions et leurs tailles. Une autre possibilité serait de quantifier l’impact des stratégies sur
le calcul de trajectoire de particules. Finalement, la comparaison des profils de pression pourrait
compléter la compréhension de l’influence des stratégies sur l’écoulement global.
Le deuxième objectif spécifique était de déterminer la précision de la LBM dans le cas de
géométries complexes. Suite aux conclusions de l’article 1, la stratégie d’imposition de
conditions aux limites EM a été choisie pour sa précision. Elle possède aussi l’avantage d’évaluer
précisément la position des frontières, ce qui est crucial pour la modélisation du mouvement
d’une partie mobile. De plus, elle permet d’étudier l’impact d’une petite variation de la
géométrie, ce qui est nécessaire à l’optimisation géométrique des procédés de mélange.
151
L’article 2 présente ainsi une étude approfondie de la précision du schéma LBM, couplé avec la
stratégie EM, dans le cas de géométries complexes avec et sans parties mobiles. Tout d’abord, la
simulation de l’écoulement de Couette cylindrique a permis d’étudier la précision du schéma
pour la norme du taux de cisaillement. Cette caractérisation est importante puisque le taux de
cisaillement est utilisé dans le calcul de la puissance consommée et interviendra dans les modèles
rhéologiques étudiés dans la dernière partie de cette thèse. Il ressort que le schéma LBM possède
une précision du deuxième ordre pour la convergence spatiale de la norme du taux de
cisaillement. Cependant, une modification mineure des conditions aux limites peut entrainer une
chute au premier ordre. Cette perte de précision globale semble être générée par des perturbations
apparaissant dans les configurations géométriques où la stratégie EM n’est pas parfaitement
définie. Cela confirme que dans le cas d’une géométrie non triviale, aucune stratégie d’imposition
de conditions aux limites n’est parfaite en LBM. Si la précision en vitesse est bien contrôlée avec
les méthodes de type interpolation ou extrapolation, des perturbations peuvent affecter la
conservation de la masse ou la précision sur le taux de cisaillement. Afin de compléter l’étude du
Couette cylindrique, deux approches possibles pour calculer la puissance ont aussi été comparées.
Il ressort que la puissance calculée à partir du couple sur l’obstacle (intégrale surfacique) est plus
précise que celle basée sur le taux de cisaillement (intégrale volumique). Cette différence
s’explique par une mauvaise évaluation du volume total du fluide, due à l’utilisation d’un
maillage structuré cartésien. En effet, lorsque la puissance est calculée à partir de l’intégrale
volumique du taux de cisaillement, l’erreur sur la puissance est dominée par l’erreur sur le
volume.
L’étude de la précision du schéma LBM a ensuite été étendue à la simulation de
l’hydrodynamique d’un système de mélange Maxblend. Ce mélangeur a été sélectionné car il est
performant sur une large gamme de nombres de Reynolds. La précision a été déterminée sur les
prédictions de consommation de puissance et de capacité de pompage. Tout d’abord, à l’aide
d’un repère Lagrangien fixé sur la pale, la configuration sans contre-pale a été simulée pour
différents nombres de Reynolds. Les valeurs sélectionnées sont représentatives des conditions
d’utilisation standards de ce système de mélange dans les régimes laminaire et transitoire. Les
résultats obtenus montrent que les consommations de puissance et les capacités de pompage sont
en accord avec les valeurs disponibles dans la littérature. De plus, la variation des performances
en fonction de l’écart entre la pale et le fond de la cuve est aussi en accord avec les études
152
précédentes. Comme la stratégie EM permet d’étudier facilement cet aspect, une étude exhaustive
de l’impact de l’espacement a été réalisée. De cette manière, il a été montré que si l’écart
standard entre la pale et la cuve est proche de l’optimal en régime laminaire, un écart deux fois
plus important serait plus performant en régime transitoire. Cette étude montre que, dans le cas de
géométries sans partie mobile, la précision du schéma LBM permet de réaliser de l’optimisation
géométrique de procédés.
Dans la dernière partie de l’article 2, la précision de la LBM a été étudiée dans le cas de
géométries avec des parties mobiles. Pour ce faire, les configurations avec et sans contre-pales
ont été simulées dans le repère naturel (eulérien du laboratoire). De cette manière, il est possible
de quantifier les erreurs associées à la modélisation de l’impact de la pale en rotation sur le
fluide. Dans un premier temps, en se basant sur les prédictions obtenues avec le repère
Lagrangien, la précision des résultats sans contre-pale a été étudiée. Il ressort que les prédictions
de consommation de puissance oscillent significativement dans le temps, ces oscillations étant
dues aux initialisations de lattices nouvellement fluides, apparaissant après le passage de la pale.
Cependant, les valeurs moyennes de consommation de puissance sont en accord avec les résultats
précédents et l’amplitude des oscillations diminue lorsque la taille de lattice est réduite. Afin de
compléter l’étude, les configurations avec contre-pales ont aussi été simulées. Dans ce cas, les
résultats montrent que l’accroissement de puissance dû à la présence des contre-pales est en
accord avec les données de la littérature. De la même manière, la variation de puissance
maximale pendant une rotation de pale est proche de la valeur obtenue dans une étude numérique
précédente utilisant la méthode des éléments finis (FEM). En conclusion, il ressort que, dans le
cas de géométries complexes avec et sans parties mobiles, la LBM permet d’obtenir des
prédictions précises de consommation de puissance et de capacité de pompage. Il est même
possible de réaliser une optimisation géométrique d’un système de mélange, ce qui peut être plus
fastidieux avec d’autres approches numériques qui nécessiteraient la génération d’un grand
nombre de maillages. Dans le cas des géométries avec parties mobiles, il serait intéressant de
compléter cette étude en utilisant la stratégie d’imposition de conditions aux limites IB-LBM. En
effet, avec cette stratégie, la modélisation du mouvement d’un obstacle se limite à déplacer les
points de contrôle et aucune initialisation de lattice n’est nécessaire. La comparaison des résultats
avec la stratégie EM permettrait d’apporter un complément d’information à la comparaison des
stratégies.
153
Le troisième objectif spécifique était d’étudier l’applicabilité de la LBM pour des rhéologies
complexes inélastiques. Dans le cas du schéma LBM, l’utilisation d’un modèle rhéologique se
limite à une modification locale du facteur de relaxation. Puisque cette intégration est aisée et ne
perturbe pas significativement la précision ou les temps de calcul, il est important de déterminer
les limites atteignables en termes de complexité rhéologique.
L’article 3 présente ainsi l’étude de l’impact de la rhéologie sur les mécanismes associés à la
création du pompage dans le système de mélange Maxblend. La stratégie de simulation est basée
sur les conclusions des deux articles précédents. Ainsi, dans le but de maximiser la précision sur
la vitesse et sur le taux de cisaillement local, les conditions aux limites ont été imposées à l’aide
de la stratégie EM. De la même manière, puisque les simulations avec des parties mobiles ont
montré un niveau de précision suffisant, il est possible de simuler la configuration avec contrepales, la plus couramment utilisée dans l’industrie. Tout d’abord, afin de définir une référence,
l’évolution de la capacité de pompage en fonction du régime d’écoulement a été caractérisée dans
le cas d’un liquide newtonien. Cette étape montre que les prédictions sont en accord avec les
données expérimentales et numériques disponibles dans la littérature. Du point de vue de
l’hydrodynamique, il ressort une différence structurelle de l’écoulement secondaire (composantes
radiales et axiales) entre le régime laminaire profond et les régimes plus élevés. Cette étape a
aussi permis de revisiter des résultats expérimentaux récents, ainsi que d’expliquer l’origine de
certaines incohérences dans les mesures réalisées à faibles nombres de Reynolds.
L’étude a ensuite été étendue à la simulation d’un liquide fortement rhéofluidifiant (indice de
puissance de 0.05) modélisé à l’aide d’un modèle de Carreau-Yasuda. Le choix du modèle
rhéologique est motivé par deux études expérimentales récentes qui montrent des patrons
d’écoulements problématiques avec ce liquide. Plus précisément, dans le régime laminaire, les
résultats expérimentaux ont souligné la présence de zones mortes volumineuses. Il est important
de souligner qu’un modèle aussi fortement rhéofluidifiant serait très difficile à simuler avec
d’autres méthodes CFD, comme celle des éléments finis. Après avoir vérifié que les prédictions
de consommation de puissance étaient en accord avec les valeurs expérimentales, l’évolution de
l’hydrodynamique en fonction du régime d’écoulement a été étudiée avec la LBM. Les résultats
ont montré que, sur une certaine plage de vitesses de rotation, des zones avec de fortes disparités
de viscosité apparente se créent. La forme de ces zones étant assez proche des observations
expérimentales, cela pourrait expliquer les patrons d’écoulements problématiques. Cela montre
154
aussi que la LBM est adéquate pour l’étude de problèmes industriels comportant des rhéologies
complexes.
Cette étude a aussi permis de déterminer les limitations associées à l’utilisation de la LBM dans
le cas de rhéologies fortement rhéofluidifiantes. En effet, comme la variation de viscosité dans le
volume est intégrée au travers du facteur de relaxation, il est nécessaire de pouvoir utiliser une
plage importante de ce paramètre. Autrement dit, le paramètre de relaxation initial, correspondant
à la viscosité maximale, doit être assez grand pour que la viscosité minimale obtenue n’entraine
pas une valeur de relaxation trop petite (i.e. trop proche de 0.5) et donc instable pour le schéma
LBM. D’un point de vue pratique, plus le modèle est rhéofluidifiant, plus le paramètre de
relaxation de référence doit être important. Cependant, l’augmentation du nombre de Reynolds
(et donc de la vitesse de rotation) nécessite de réduire le facteur de relaxation pour conserver la
stabilité numérique et la précision de la méthode. Il ressort que l’utilisation de modèles fortement
rhéofluidifiants entraine une limitation sur les nombres de Reynolds simulables. Par exemple,
dans le cadre de l’article 3, avec une taille de lattice de 1.77 mm, il n’a pas été possible de
dépasser une valeur de 60 pour le nombre de Reynolds généralisé. L’avantage est que, lorsque les
simulations sont possibles, les temps de calcul restent semblables à ceux des cas newtoniens.
Afin de dépasser les limitations sur les nombres de Reynolds simulables, il serait intéressant de
tester d’autres modèles d’opérateur de collision. Par exemple, il est possible que l’utilisation d’un
schéma avec plusieurs facteurs de relaxation (comme la MRT) permette d’améliorer la stabilité
des calculs pour de faibles valeurs du facteur de relaxation (proches de 0.5).
155
CONCLUSION ET RECOMMANDATIONS
L’objectif principal de cette thèse était de développer une méthode de Boltzmann sur réseau
(LBM) efficace pour la simulation des écoulements visqueux en cuves agitées. Pour ce faire, trois
articles ont été présentés.
La simulation des procédés industriels de mélange visqueux est caractérisée par la difficulté à
traiter la complexité de la géométrie, à intégrer la rhéologie de la mixture ainsi qu’à modéliser
l’interaction entre les différentes phases. La plupart des travaux de recherche dans ce domaine se
concentrent sur la modélisation, avec des méthodes classiques, de l’impact des pales sur le fluide
et sur l’intégration des modèles rhéologiques. Comme les temps de calcul sont très importants, la
majorité des études concernent des écoulements monophasiques dont les propriétés rhéologiques
ne sont pas fortement non newtoniennes. Afin de dépasser ces limitations, il est intéressant
d’étudier les possibilités offertes par d’autres méthodes numériques. Le travail présenté dans cette
thèse a permis d’apporter une contribution à ce champ de recherche.
Tout d’abord, la comparaison de trois stratégies d’imposition de conditions aux limites a permis
d’étudier leur impact sur le précision du schéma LBM ainsi que leur applicabilité à la simulation
de systèmes de mélange. L’étude de l’écoulement de Couette cylindrique a permis de montrer
que seule une stratégie basée sur l’interpolation ou l’extrapolation est capable de conserver la
précision naturelle du schéma LBM pour la vitesse. À l’inverse, les autres approches génèrent des
oscillations sur les vitesses aux alentours des frontières. L’étude des paramètres numériques
associés à chaque stratégie a montré que seule la stratégie basée sur une condition aux limites
immergées est sensible aux paramètres numériques. En effet, la disposition des points de
contrôle, utilisés pour modéliser l’impact de la partie en rotation, affecte largement l’erreur sur la
solution numérique. La comparaison a ensuite été étendue à la simulation de systèmes de
mélange comprenant une ancre et un Paravisc. Il ressort que les prédictions de consommation de
puissance et de capacité de pompage sont similaires avec l’approche la plus simple, la méthode
du rebond modifié (MBB), ou avec l’utilisation d’une méthode d’extrapolation. Dans le cas de la
stratégie de conditions aux limites immergées, les résultats sont proches mais présentent une
variabilité conséquente en fonction de la taille de lattice. Il ressort que, dans le cas des
mélangeurs raclants, cette stratégie ne permet pas d’obtenir la même précision, à cause de
156
l’apparition de perturbations sur la vitesse dans la zone de la pale. En conclusion, les stratégies
d’interpolation ou d’extrapolation sont les mieux adaptées car elles permettent de minimiser les
erreurs et prendre en compte de petites variations de la géométrie.
La deuxième partie de la thèse présentait une étude plus approfondie de la précision du schéma
LBM, couplé avec la stratégie d’extrapolation, dans le cas de géométries avec et sans parties
mobiles. L’étude de la précision sur la norme du taux de cisaillement a montré que cette grandeur
est plus sensible que la vitesse à la définition des conditions aux limites, confirmant que même
les stratégies les plus précises d’imposition de conditions aux limites perturbent le schéma LBM.
De plus, les différentes méthodes disponibles pour calculer la puissance présentent des précisions
différentes en fonction de la stratégie de simulation, c’est-à-dire avec ou sans parties mobiles.
D’un point de vue plus général, la simulation du système de mélange Maxblend, dans les
configurations avec et sans contre-pales, a permis de conclure que la LBM peut être efficacement
utilisée pour simuler ce type de système de mélange. En effet, dans le cas d’une géométrie sans
partie mobile (utilisation du repère lagrangien), la méthode a pu être utilisée pour étudier l’impact
de l’espacement entre la pale et la cuve, montrant ainsi la possibilité de faire de l’optimisation
géométrique avec cette approche. Dans le cas des géométries avec parties mobiles, les prédictions
de consommations de puissance présentent des valeurs moyennes en accord avec les données de
la littérature. Notons que l’initialisation des lattices nouvellement fluides, requise après le
passage d’un objet, perturbe le champ de vitesse autour de la pale. Cependant, l’amplitude de ces
oscillations diminue avec la taille de lattices.
La dernière partie de la thèse traite de l’applicabilité de la LBM pour des rhéologies complexes
inélastiques. Comme l’intégration d’un modèle rhéologique dans le schéma LBM peut se faire
d’une manière simple, qui n’affecte pas significativement la précision ou les temps de calcul, les
limites atteignables avec un liquide fortement rhéofluidifiant (indice de puissance de 0.05) ont été
étudiées. Pour ce faire, la caractérisation de l’écoulement secondaire (composantes axiales et
radiales de l’écoulement) a permis de souligner les limites sur les nombres de Reynolds
simulables pour une taille de lattice donnée. En effet, comme l’introduction du modèle de
rhéologie se fait à travers le paramètre de relaxation, celui-ci doit représenter la variation totale
de viscosité. Cependant, pour des raisons de stabilité, l’augmentation de la vitesse de rotation
nécessite une réduction du paramètre de relaxation, ce qui limite la plage de nombres de
Reynolds simulables pour une taille de lattice donnée. Malgré ces limitations, la méthode LBM a
157
pu être utilisée pour investiguer un problème industriel relatif à l’utilisation d’une rhéologie
complexe.
En conclusion, la méthode de Boltzmann sur réseau a été utilisée avec succès pour la simulation
de l’agitation d’écoulements monophasiques visqueux dans des systèmes de mélange. Les
capacités et les limites pour le traitement de la complexité de la géométrie et de la rhéologie ont
été caractérisées. S’il est certain que cette méthode ne va pas remplacer les méthodes des
éléments et des volumes finis, elle peut être utilisée à bon escient dans certaines configurations.
Par exemple, elle constitue une piste sérieuse pour tenter de simuler des écoulements
polyphasiques ou à rhéologies complexes à l’échelle industrielle. Afin de continuer ce travail de
recherche, les recommandations suivantes sont proposées :
1. Le raffinement local de maillage étant particulièrement adapté aux géométries des
mélangeurs raclants (présence d’un entrefer), son application devrait permettre des gains
de précision et de temps de calcul conséquents. Une automatisation de ce processus serait
requise pour le traitement des parties mobiles.
2. Il serait intéressant d’étendre la comparaison des stratégies d’imposition de conditions
aux limites à des écoulements comprenant des parties mobiles. Il est probable que les
avantages des approches basés sur des conditions aux limites immergées ressortent plus
clairement.
3. Il serait intéressant de déterminer les limites atteignables pour d’autres types de
phénomènes rhéologiques. Par exemple, il faudrait quantifier l’impact de l’introduction
des effets d’élasticité ou de thixotropie sur les possibilités du schéma LBM. Comme cette
étape nécessite l’ajout d’une équation supplémentaire, il serait pertinent de comparer les
différentes approches.
4. Puisque les méthodes de types particules discrètes (DPM) ont été appliquées avec succès
à la LBM, il serait intéressant de voir dans quelles mesures cette méthode permet de
recréer les viscosités apparentes de suspensions. Une attention particulière devra être
portée sur le couplage solide-liquide.
158
BIBLIOGRAPHIE
[1]
D. Vidal, C. Ridgway, G. Pianet, J. Schoelkopf, R. Roy, and F. Bertrand, "Effect of
particle size distribution and packing compression on fluid permeability as predicted by
lattice-Boltzmann simulations," Computers and Chemical Engineering, vol. 33, pp. 256266, 2009.
[2]
D. Vidal, R. Roy, and F. Bertrand, "A parallel workload balanced and memory efficient
lattice-Boltzmann algorithm with single unit BGK relaxation time for laminar Newtonian
flows," Computers and Fluids, vol. 39, pp. 1411-1423, 2010.
[3]
D. Vidal, R. Roy, and F. Bertrand, "On improving the performance of large parallel lattice
Boltzmann flow simulations in heterogeneous porous media," Computers and Fluids, vol.
39, pp. 324-37, 2010.
[4]
R. Hilfer, "Local Porosity Theory and Stochastic Reconstruction for Porous Media." vol.
554, K. Mecke and D. Stoyan, Eds., ed: Springer Berlin / Heidelberg, 2000, pp. 203-241.
[5]
K. Vafai, Handbook of porous media: Taylor & Francis, 2005.
[6]
R. E. Hayes, A. Fadic, J. Mmbaga, and A. Najafi, "CFD modelling of the automotive
catalytic converter," Catalysis Today, vol. 188, pp. 94-105, 2012.
[7]
F. Bertrand, C. Devals, D. Vidal, C. S. d. Preval, and R. E. Hayes, "Towards the
simulation of the catalytic monolith converter using discrete channel-scale models,"
Catalysis Today, 2012.
[8]
I. Demirdžić and M. Perić, "Finite volume method for prediction of fluid flow in
arbitrarily shaped domains with moving boundaries," International Journal for Numerical
Methods in Fluids, vol. 10, pp. 771-790, 1990.
[9]
C. S. Peskin, "Numerical analysis of blood flow in the heart," Journal of Computational
Physics, vol. 25, pp. 220-252, 1977.
[10]
R. Glowinski, T.-W. Pan, and J. Periaux, "Fictitious domain method for Dirichlet problem
and applications," Computer Methods in Applied Mechanics and Engineering, vol. 111,
pp. 283-303, 1994.
[11]
R. B. Bird, Armstrong, R.C., Hassager, O., Dynamics of polymer liquids vol. 1 2nd ed.:
Wiley, New York, 1987.
[12]
F. Bertrand, P. A. Tanguy, E. Brito De La Fuente, and P. Carreau, "Numerical modeling
of the mixing flow of second-order fluids with helical ribbon impellers," Computer
Methods in Applied Mechanics and Engineering, vol. 180, pp. 267-280, 1999.
[13]
C. Devals, M. Heniche, K. Takenaka, and P. A. Tanguy, "CFD analysis of several design
parameters affecting the performance of the Maxblend impeller," Computers and
Chemical Engineering, vol. 32, pp. 1831-1841, 2008.
[14]
O. P. Malaspinas, "Lattice-Boltzmann method for the simulation of viscoelastic fluid
flows," Ph.D., PROGRAMME DOCTORAL EN MÉCANIQUE, École Polytechnique
Fédérale de Lausanne, 2009.
159
[15]
J. Boyd, J. Buick, and S. Green, "A second-order accurate lattice Boltzmann nonNewtonian flow model," Journal of Physics A (Mathematical and General), vol. 39, pp.
14241-7, 2006.
[16]
J. M. Buick, "Lattice Boltzmann simulation of power-law fluid flow in the mixing section
of a single-screw extruder," Chemical Engineering Science, vol. 64, pp. 52-58, 2009.
[17]
J. Boyd, J. M. Buick, and S. Green, "Analysis of the Casson and Carreau-Yasuda nonNewtonian blood models in steady and oscillatory flows using the lattice Boltzmann
method," Physics of Fluids, vol. 19, 2007.
[18]
G. Pingen and K. Maute, "Optimal design for non-Newtonian flows using a topology
optimization approach," Computers and Mathematics with Applications, vol. 59, pp.
2340-2350, 2010.
[19]
A. Lefebvre and B. Maury, "Apparent viscosity of a mixture of a Newtonian fluid and
interacting particles," Comptes Rendus Mécanique, vol. 333, pp. 923-933, 2005.
[20]
M. A. Van Der Hoef, M. Van Sint Annaland, and J. A. M. Kuipers, "Computational fluid
dynamics for dense gas-solid fluidized beds: A multi-scale modeling strategy," in
ISCRE18, 2004, pp. 5157-5165.
[21]
R. J. Phillips, R. C. Armstrong, R. A. Brown, A. L. Graham, and J. R. Abbott,
"Constitutive equation for concentrated suspensions that accounts for shear-induced
particle migration," Physics of Fluids A (Fluid Dynamics), vol. 4, pp. 30-30, 1992.
[22]
L. Fradette, P. A. Tanguy, F. Bertrand, F. Thibault, J.-B. Ritz, and E. Giraud, "CFD
phenomenological model of solid-liquid mixing in stirred vessels," Computers and
Chemical Engineering, vol. 31, pp. 334-345, 2007.
[23]
J. A. M. Kuipers, K. J. V. Duin, P. H. Van Beckum, and W. P. M. Van Swaali,
"Numerical model of gas-fluidized beds," Chemical Engineering Science, vol. 47, pp.
1913-1924, 1992.
[24]
R. S. Oey, R. F. Mudde, L. M. Portela, and H. E. A. Van Den Akker, "Simulation of a
slurry airlift using a two-fluid model," Chemical Engineering Science, vol. 56, pp. 673681, 2001.
[25]
G. R. Kasat, A. R. Khopkar, V. V. Ranade, and A. B. Pandit, "CFD simulation of liquidphase mixing in solid–liquid stirred reactor," Chemical Engineering Science, vol. 63, pp.
3877-3885, 2008.
[26]
J. J. Derksen, "Numerical Simulation of Solids Suspension in a Stirred Tank," AIChE
Journal, vol. 49, pp. 2700-2714, 2003.
[27]
J. J. Derksen, "Solid particle mobility in agitated bingham liquids," Industrial and
Engineering Chemistry Research, vol. 48, pp. 2266-2274, 2009.
[28]
G. Pianet, F. Bertrand, D. Vidal, and B. Mallet, "Discrete element method-based models
for the consolidation of particle packings in paper-coating applications," Asia-Pacific
Journal of Chemical Engineering, vol. 6, pp. 44-54, 2011.
[29]
Y. C. Zhou, B. D. Wright, R. Y. Yang, B. H. Xu, and A. B. Yu, "Rolling friction in the
dynamic simulation of sandpile formation," Physica A: Statistical Mechanics and its
Applications, vol. 269, pp. 536-553, 1999.
160
[30]
F. Bertrand, L. A. Leclaire, and G. Levecque, "DEM-based models for the mixing of
granular materials," Chemical Engineering Science, vol. 60, pp. 2517-2531, 2005.
[31]
A. B. Metzner and R. E. Otto, "Agitation of non-newtonian fluids," Chemical
Engineering Progress, vol. 3, pp. 3-10, 1957.
[32]
F. Cabaret, S. Bonnot, L. Fradette, and P. A. Tanguy, "Mixing time analysis using
colorimetric methods and image processing," Industrial and Engineering Chemistry
Research, vol. 46, pp. 5032-5042, 2007.
[33]
Y. Guntzburger, L. Fradette, M. Farhat, M. Heniche, P. A. Tanguy, and K. Takenaka,
"Effect of the geometry on the performance of the Maxblend impeller with viscous
newtonian fluids," Asia-Pacific Journal of Chemical Engineering, vol. 4, pp. 528-536,
2009.
[34]
K. Yamamoto, A. Katumi, A. Tarumoto, K. NISHI, M. Kaminoyama, and M. Kamiwano,
"Development and evaluation of large-scale impeller generating strong circulation flow
suitable for wide viscosity range in reactor with cooling coil," Journal of Chem. Eng.
Jap., vol. 31, 1998.
[35]
Y. Guntzburger, A. Fontaine, L. Fradette, and F. Bertrand, "An experimental method to
evaluate global pumping in a mixing system: application to the Maxblend™ for
Newtonian and non-Newtonian fluids," In press, Accepted Manuscript, Chemical
Engineering Journal.
[36]
A. Iranshahi, C. Devals, M. Heniche, L. Fradette, P. A. Tanguy, and K. Takenaka,
"Hydrodynamics characterization of the Maxblend impeller," Chemical Engineering
Science, vol. 62, pp. 3641-3653, 2007.
[37]
P. A. Tanguy, R. Lacroix, F. Bertrand, L. Choplin, and E. B. de la Fuente, "Finite element
analysis of viscous mixing with a helical ribbon-screw impeller," AIChE Journal, vol. 38,
pp. 939-944, 1992.
[38]
A. Brucato, M. Ciofalo, F. Grisafi, and G. Micale, "Numerical prediction of flow fields in
baffled stirred vessels: A comparison of alternative modelling approaches," Chemical
Engineering Science, vol. 53, pp. 3653-3684, 1998.
[39]
O. Estruch, O. Lehmkuhl, R. Borrell, C. D. Pérez Segarra, and A. Oliva, "A parallel radial
basis function interpolation method for unstructured dynamic meshes," Computers &
Fluids, 2012.
[40]
F. Bertrand, P. A. Tanguy, and F. Thibault, "A three-dimensional fictitious domain
method for incompressible fluid flow problems," International Journal for Numerical
Methods in Fluids, vol. 25, pp. 719-36, 1997.
[41]
A. Iranshahi, M. Heniche, F. Bertrand, and P. A. Tanguy, "Numerical investigation of the
mixing efficiency of the Ekato Paravisc impeller," Chemical Engineering Science, vol.
61, pp. 2609-2617, 2006.
[42]
K. S. Sujatha and M. F. Webster, "Modelling three-dimensional rotating flows in
cylindrical-shaped vessels," International Journal for Numerical Methods in Fluids, vol.
43, pp. 1067-1079, 2003.
161
[43]
W. Yao, M. Mishima, and K. Takahashi, "Numerical investigation on dispersive mixing
characteristics of MAXBLEND and double helical ribbons," Chemical Engineering
Journal, vol. 84, pp. 565-571, 2001.
[44]
B. Coesnon, M. Heniche, C. Devals, F. Bertrand, and P. A. Tanguy, "A fast and robust
fictitious domain method for modelling viscous flows in complex mixers: The example of
propellant make-down," International Journal for Numerical Methods in Fluids, vol. 58,
pp. 427-449, 2008.
[45]
T. Jongen, "Characterization of batch mixers using numerical flow simulations," AIChE
Journal, vol. 46, pp. 2140-2150, 2000.
[46]
C. A. Rivera, M. Heniche, K. Takenaka, and P. A. Tanguy, "Finite element modeling of
the laminar and transition flow of the Superblend dual shaft coaxial mixer on parallel
computers," Chemical Engineering Science, vol. 64, pp. 4442-56, 2009.
[47]
F. Bertrand, P. A. Tanguy, and E. Brito-De La Fuente, "A New Perspective for the
Mixing of Yield Stress Fluids with Anchor Impellers," JOURNAL OF CHEMICAL
ENGINEERING OF JAPAN, vol. 29, pp. 51-58, 1996.
[48]
Z. Jaworski, W. Bujalski, N. Otomo, and A. W. Nienow, "CFD study of homogenization
with dual rushton turbines-comparison with experimental results. Part I: Initial studies," in
Fluid Mixing, 2000, pp. 327-333.
[49]
G. L. Lane, M. P. Schwarz, and G. M. Evans, "Predicting gas-liquid flow in a
mechanically stirred tank," Applied Mathematical Modelling, vol. 26, pp. 223-235, 2002.
[50]
B. Basara, A. Alajbegovic, and D. Beader, "Simulation of single- and two-phase flows on
sliding unstructured meshes using finite volume method," International Journal for
Numerical Methods in Fluids, vol. 45, pp. 1137-1159, 2004.
[51]
K. Yapici, B. Karasozen, M. Schafer, and Y. Uludag, "Numerical investigation of the
effect of the Rushton type turbine design factors on agitated tank flow characteristics,"
Chemical Engineering and Processing: Process Intensification, vol. 47, pp. 1346-1355,
2008.
[52]
L. Rudolph, M. Schafer, V. Atiemo-Obeng, and M. Kraume, "Experimental and
numerical analysis of power consumption for mixing of high viscosity fluids with a coaxial mixer," 165-189 Railway Terrace, Rugby, Warks., CV21 3HQ, United Kingdom,
2007, pp. 568-575.
[53]
A. J. C. Ladd and R. Verberg, "Lattice-Boltzmann simulations of particle-fluid
suspensions," Journal of Statistical Physics, vol. 104, pp. 1191-251, 2001.
[54]
J. J. Derksen and Prashant, "Simulations of complex flow of thixotropic liquids," Journal
of Non-Newtonian Fluid Mechanics, vol. 160, pp. 65-75, 2009.
[55]
D. Kandhai, "Large scale lattice-Boltzmann simulations-Computational methods and
applications," Ph.D., Amsterdam university, 1999.
[56]
J. Latt, "Hydrodynamic Limit of Lattice Boltzmann Equations," Ph.D., Faculté des
sciences, Université de Genève, 2007.
[57]
K. Mattila, "Implementation Techniques for the Lattice Boltzmann Method," Ph.D.,
University of Jyvaskyla, 2010.
162
[58]
S. Succi, The lattice Boltzmann equation for fluid dynamics and beyond. Oxford:
Clarendon Press, 2001.
[59]
X. Shan, X.-F. Yuan, and H. Chen, "Kinetic theory representation of hydrodynamics: A
way beyond the Navier-Stokes equation," Journal of Fluid Mechanics, vol. 550, pp. 413441, 2006.
[60]
P. L. Bhatnagar, E. P. Gross, and M. Krook, "A Model for Collision Processes in Gases. I.
Small Amplitude Processes in Charged and Neutral One-Component Systems," Physical
Review, vol. 94, p. 511, 1954.
[61]
S. Chen and G. D. Doolen, "Lattice Boltzmann method for fluid flows," Annual Review of
Fluid Mechanics, vol. 30, pp. 329-364, 1998.
[62]
T. Kruger, F. Varnik, and D. Raabe, "Shear stress in lattice Boltzmann simulations,"
Physical Review E - Statistical, Nonlinear, and Soft Matter Physics, vol. 79, 2009.
[63]
Y. H. Qian, D. D'Humieres, and P. Lallemand, "Lattice BGK models for Navier-Stokes
equation," Europhysics Letters, vol. 17, pp. 479-84, 1992.
[64]
D. Kandhai, A. Koponen, A. Hoekstra, M. Kataja, J. Timonen, and P. M. A. Sloot,
"Implementation aspects of 3D lattice-BGK: boundaries, accuracy, and a new fast
relaxation method," J. Comput. Phys., vol. 150, pp. 482-501, 1999.
[65]
Z. Guo, C. Zheng, and B. Shi, "Discrete lattice effects on the forcing term in the lattice
Boltzmann method," Physical Review E. Statistical Physics, Plasmas, Fluids, and Related
Interdisciplinary Topics, vol. 65, pp. 46308/1-46308/6, 2002.
[66]
M. Junk, A. Klar, and L.-S. Luo, "Asymptotic analysis of the lattice Boltzmann equation,"
Journal of Computational Physics, vol. 210, pp. 676-704, 2005.
[67]
M. Junk and Z. X. Yang, "Convergence of lattice Boltzmann methods for Navier-Stokes
flows in periodic and bounded domains," Numerische Mathematik, vol. 112, pp. 65-87,
Mar 2009.
[68]
X. Shan and H. Chen, "Simulation of nonideal gases and liquid-gas phase transitions by
the lattice Boltzmann equation," Physical Review E, vol. 49, pp. 2941-2948, 1994.
[69]
M. R. Swift, W. R. Osborn, and J. M. Yeomans, "Lattice Boltzmann Simulation of
Nonideal Fluids," Physical Review Letters, vol. 75, pp. 830-833, 1995.
[70]
S. Leclaire, M. Reggio, and J.-Y. Trépanier, "Isotropic color gradient for simulating very
high-density ratios with a two-phase flow lattice Boltzmann model," Computers &
Fluids, vol. 48, pp. 98-112, 2011.
[71]
D. P. Ziegler, "Boundary conditions for lattice Boltzmann simulations," Journal of
Statistical Physics, vol. 71, pp. 1171-1177, 1993.
[72]
A. J. C. Ladd, "Numerical simulations of particulate suspensions via a discretized
Boltzmann equation. Part 1. Theoretical foundation," Journal of Fluid Mechanics, vol.
271, pp. 285-309, 1994.
[73]
X. He, Q. Zou, L.-S. Luo, and M. Dembo, "Analytic solutions of simple flows and
analysis of nonslip boundary conditions for the lattice Boltzmann BGK model," Journal
of Statistical Physics, vol. 87, pp. 115-136, 1997.
163
[74]
Y. K. Suh, J. Kang, and S. Kang, "Assessment of algorithms for the no-slip boundary
condition in the lattice Boltzmann equation of BGK model," International Journal for
Numerical Methods in Fluids, vol. 58, pp. 1353-1378, 2008.
[75]
S. Chen, D. Martinez, and R. Mei, "On boundary conditions in lattice Boltzmann
methods," Physics of Fluids, vol. 8, pp. 2527-2527, 1996.
[76]
R. Mei, L.-S. Luo, and W. Shyy, "An accurate curved boundary treatment in the lattice
Boltzmann method," Journal of Computational Physics, vol. 155, pp. 307-30, 1999.
[77]
M. Bouzidi, M. Firdaouss, and P. Lallemand, "Momentum transfer of a Boltzmann-lattice
fluid with boundaries," Physics of Fluids, vol. 13, pp. 3452-3459, 2001.
[78]
Z. Guo, C. Zheng, and B. Shi, "An extrapolation method for boundary conditions in
lattice Boltzmann method," Physics of Fluids, vol. 14, pp. 2007-10, 2002.
[79]
D. Yu, R. Mei, L.-S. Luo, and W. Shyy, "Viscous flow computations with the method of
lattice Boltzmann equation," Progress in Aerospace Sciences, vol. 39, pp. 329-367, 2003.
[80]
J. J. Derksen, J. L. Kooman, and H. E. A. van den Akker, "Parallel fluid flow simulations
by means of a lattice-Boltzmann scheme," High-Performance Computing and
Networking, vol. 1225, pp. 524-530, 1997.
[81]
Z. G. Feng and E. E. Michaelides, "The immersed boundary-lattice Boltzmann method for
solving fluid-particles interaction problems," Journal of Computational Physics, vol. 195,
pp. 602-28, 2004.
[82]
A. Dupuis, P. Chatelain, and P. Koumoutsakos, "An immersed boundary-latticeBoltzmann method for the simulation of the flow past an impulsively started cylinder,"
Journal of Computational Physics, vol. 227, pp. 4486-98, 2008.
[83]
J. Wu and C. Shu, "Implicit velocity correction-based immersed boundary-lattice
Boltzmann method and its applications," Journal of Computational Physics, vol. 228, pp.
1963-79, 2009.
[84]
J. Latt, B. Chopard, O. Malaspinas, M. Deville, and A. Michler, "Straight velocity
boundaries in the lattice Boltzmann method," Physical Review E, vol. 77, p. 056703,
2008.
[85]
C. S. Peskin, "The immersed boundary method," Acta Numerica, vol. 11, pp. 479-517,
2002.
[86]
J. Derksen and H. E. A. Van Den Akker, "Large eddy simulations on the flow driven by a
Rushton turbine," AIChE Journal, vol. 45, pp. 209-221, 1999.
[87]
P. Lallemand and L. Li-Shi, "Lattice Boltzmann method for moving boundaries," Journal
of Computational Physics, vol. 184, pp. 406-21, 2003.
[88]
A. Caiazzo, "Analysis of lattice Boltzmann nodes initialisation in moving boundary
problems," Progress in Computational Fluid Dynamics, vol. 8, pp. 3-10, 2008.
[89]
R. Mei, D. Yu, W. Shyy, and L.-S. Luo, "Force evaluation in the lattice Boltzmann
method involving curved geometry," Physical Review E - Statistical, Nonlinear, and Soft
Matter Physics, vol. 65, pp. 041203/1-041203/14, 2002.
164
[90]
M. Ohta, T. Nakamura, Y. Yoshida, and Y. Matsukuma, "Lattice Boltzmann simulations
of viscoplastic fluid flows through complex flow channels," Journal of Non-Newtonian
Fluid Mechanics, vol. 166, pp. 404-412, 2011.
[91]
N. Goyal and J. J. Derksen, "Direct simulations of spherical particles sedimenting in
viscoelastic fluids," Journal of Non-Newtonian Fluid Mechanics, vol. 183–184, pp. 1-13,
2012.
[92]
J. J. J. Gillissen and H. E. A. Van den Akker, "Direct numerical simulation of the
turbulent flow in a baffled tank driven by a Rushton turbine," AIChE Journal, pp. n/a-n/a,
2012.
[93]
H. Hartmann, J. J. Derksen, C. Montavon, J. Pearson, I. S. Hamill, and H. E. A. van den
Akker, "Assessment of large eddy and RANS stirred tank simulations by means of LDA,"
Chemical Engineering Science, vol. 59, pp. 2419-2432, 2004.
[94]
J. J. Derksen, "Agitation and mobilization of thixotropic liquids," AIChE Journal, vol. 56,
pp. 2236-2247, 2010.
[95]
J. J. Derksen, "Blending of miscible liquids with different densities starting from a
stratified state," Computers and Fluids, vol. 50, pp. 35-45, 2011.
[96]
J. J. Derksen, "Direct simulations of mixing of liquids with density and viscosity
differences," Industrial and Engineering Chemistry Research, vol. 51, pp. 6948-6957,
2012.
[97]
D. Guha, P. A. Ramachandran, M. P. Dudukovic, and J. J. Derksen, "Evaluation of large
Eddy simulation and Euler-Euler CFD models for solids flow dynamics in a stirred tank
reactor," AIChE Journal, vol. 54, pp. 766-778, 2008.
[98]
H. Hartmann, J. J. Derksen, and H. E. A. Van Den Akker, "Numerical simulation of a
dissolution process in a stirred tank reactor," Chemical Engineering Science, vol. 61, pp.
3025-3032, 2006.
[99]
R. Sungkorn, J. J. Derksen, and J. G. Khinast, "Modeling of aerated stirred tanks with
shear-thinning power law liquids," International Journal of Heat and Fluid Flow, vol. 36,
pp. 153-166, 2012.
[100] R. Sungkorn, J. J. Derksen, and J. G. Khinast, "Euler-Lagrange modeling of a gas-liquid
stirred reactor with consideration of bubble breakage and coalescence," AIChE Journal,
vol. 58, pp. 1356-1370, 2012.
[101] L. Misici and M. Felici, "Mixing in a stirred vessel: a parallel implementation with lattice
Boltzmann colored particles model," TASK Quarterly, vol. 9, pp. 449-59, 2005.
[102] P. E. Arratia, J. Kukura, J. Lacombe, and F. J. Muzzio, "Mixing of shear-thinning fluids
with yield stress in stirred tanks," AIChE Journal, vol. 52, pp. 2310-2322, 2006.
[103] C. Rivera, S. Foucault, M. Heniche, T. Espinosa-Solares, and P. A. Tanguy, "Mixing
analysis in a coaxial mixer," Chemical Engineering Science, vol. 61, pp. 2895-2907,
2006.
[104] H. M. Blackburn, J. R. Elston, D. A. Niclasen, M. Rudman, and J. Wu, "A hybrid method
for simulation of axial flow impeller driven mixing vessels," Applied Mathematical
Modelling, vol. 24, pp. 795-805, 2000.
165
[105] F. Thibault and P. A. Tanguy, "Power-draw analysis of a coaxial mixer with Newtonian
and non-Newtonian fluids in the laminar regime," Chemical Engineering Science, vol. 57,
pp. 3861-3872, 2002.
[106] J. M. Zalc, M. M. Alvarez, F. J. Muzzio, and B. E. Arik, "Extensive validation of
computed laminar flow in a stirred tank with three Rushton turbines," AIChE Journal, vol.
47, pp. 2144-2154, 2001.
[107] J. M. Zalc, E. S. Szalai, M. M. Alvarez, and F. J. Muzzio, "Using CFD to understand
chaotic mixing in laminar stirred tanks," AIChE Journal, vol. 48, pp. 2124-2134, 2002.
[108] R. K. Connelly and J. Valenti-Jordan, "Mixing analysis of a Newtonian fluid in a 3D
planetary pin mixer," Chemical Engineering Research and Design, vol. 86, pp. 14341440, 2008.
[109] T. Avalosse, "Numerical simulation of distributive
Macromolecular Symposia, vol. 112, pp. 91-98, Dec 1996.
mixing
in
3-D
flows,"
[110] A. Nabovati, E. W. Llewellin, and A. C. M. Sousa, "A general model for the permeability
of fibrous porous media based on fluid flow simulations using the lattice Boltzmann
method," Composites Part A: Applied Science and Manufacturing, vol. 40, pp. 860-869,
2009.
[111] Z.-G. Feng and E. E. Michaelides, "Proteus: a direct forcing method in the simulations of
particulate flows," Journal of Computational Physics, vol. 202, pp. 20-51, 2005.
[112] Z. Chai, B. Shi, J. Lu, and Z. Guo, "Non-Darcy flow in disordered porous media: A lattice
Boltzmann study," Computers and Fluids, vol. 39, pp. 2069-2077, 2010.
[113] K. N. Premnath, J.-C. Nave, and S. Banerjee, "Computation of multiphase flows with
Lattice Boltzmann methods," in 2005 ASME International Mechanical Engineering
Congress and Exposition, IMECE 2005, November 5, 2005 - November 11, 2005,
Orlando, FL, United states, 2005, pp. 403-420.
[114] H. Huang, L. Wang, and X.-y. Lu, "Evaluation of three lattice Boltzmann models for
multiphase flows in porous media," 2010.
[115] S. P. Sullivan, L. F. Gladden, and M. L. Johns, "Simulation of power-law fluid flow
through porous media using lattice Boltzmann techniques," Journal of Non-Newtonian
Fluid Mechanics, vol. 133, pp. 91-98, 2006.
[116] G. Mayer and G. Hazi, "Direct numerical and large eddy simulation of longitudinal flow
along triangular array of rods using the lattice Boltzmann method," Mathematics and
Computers in Simulation, vol. 72, pp. 173-178, 2006.
[117] J. J. Derksen, K. Kontomaris, J. B. McLaughlin, and H. E. A. Van den Akker, "Largeeddy simulations of single-phase flow dynamics and mixing in an industrial crystallizer,"
Chemical Engineering Research and Design, vol. 85, pp. 169-179, 2007.
[118] M. Junk and Z. Yang, "One-point boundary condition for the lattice Boltzmann method,"
Physical Review E - Statistical, Nonlinear, and Soft Matter Physics, vol. 72, pp. 1-8,
2005.
166
[119] U. Frisch, D. d'Humieres, B. Hasslacher, P. Lallemand, Y. Pomeau, and J. P. Rivet,
"Lattice gas hydrodynamics in two and three dimensions," in Workshop on Large
Nonlinear Systems, 27-29 Oct. 1986, USA, 1987, pp. 649-707.
[120] R. Benzi, S. Succi, and M. Vergassola, "The Lattice Boltzmann-Equation - Theory and
Applications," Physics Reports-Review Section of Physics Letters, vol. 222, pp. 145-197,
Dec 1992.
[121] X. He and L.-S. Luo, "Theory of the lattice Boltzmann method: from the Boltzmann
equation to the lattice Boltzmann equation," Physical Review E (Statistical Physics,
Plasmas, Fluids, and Related Interdisciplinary Topics), vol. 56, pp. 6811-17, 1997.
[122] D. D'Humieres, I. Ginzburg, M. Krafczyk, P. Lallemand, and L. Li-Shi, "Multiplerelaxation-time lattice Boltzmann models in three dimensions," in 10th International
Conference on Discrete Simulation of Fluid Dynamics, 2-6 July 2001, UK, 2002, pp. 43751.
[123] J. Latt and B. Chopard, "Lattice Boltzmann method with regularized pre-collision
distribution functions," Mathematics and Computers in Simulation, vol. 72, pp. 165-168,
2006.
[124] C. Pan, L.-S. Luo, and C. T. Miller, "An evaluation of lattice Boltzmann schemes for
porous medium flow simulation," Computers & Fluids, vol. 35, pp. 898-909, 2006.
[125] D. R. Noble, C. Shiyi, J. G. Georgiadis, and R. O. Buckius, "A consistent hydrodynamic
boundary condition for the lattice Boltzmann method," Physics of Fluids, vol. 7, pp. 2039, 1995.
[126] H. Chen, C. Teixeira, and K. Molvig, "Realization of fluid boundary conditions via
discrete Boltzmann dynamics," in Proceedings of the 7th International Conference on the
Discrete Simulation of Fluids, 14-18 July 1998, Singapore, 1998, pp. 1281-92.
[127] M. Rohde, J. J. Derksen, and H. E. A. Van Den Akker, "Volumetric method for
calculating the flow around moving objects in lattice-Boltzmann schemes," Physical
Review E - Statistical, Nonlinear, and Soft Matter Physics, vol. 65, pp. 056701/1056701/11, 2002.
[128] Y. Li, R. Shock, R. Zhang, and H. Chen, "Numerical study of flow past an impulsively
started cylinder by the lattice-Boltzmann method," Journal of Fluid Mechanics, vol. 519,
pp. 273-300, 2004.
[129] J. Lee and S. Lee, "Boundary treatment for the lattice Boltzmann method using adaptive
relaxation times," Computers and Fluids, vol. 39, pp. 900-909, 2010.
[130] A. M. Roma, C. S. Peskin, and M. J. Berger, "An Adaptive Version of the Immersed
Boundary Method," Journal of Computational Physics, vol. 153, pp. 509-534, 1999.
[131] P. A. Tanguy, F. Thibault, E. Brito-De La Fuente, T. Espinosa-Solares, and A. Tecante,
"Mixing performance induced by coaxial flat blade-helical ribbon impellers rotating at
different speeds," Chemical Engineering Science, vol. 52, pp. 1733-1741, 1997.
[132] S. Foucault, G. Ascanio, and P. A. Tanguy, "Power characteristics in coaxial mixing:
Newtonian and non-newtonian fluids," Industrial and Engineering Chemistry Research,
vol. 44, pp. 5036-5043, 2005.
167
[133] G. Delaplace, J. C. Leuliet, and G. Ronse, "Power Requirement when Mixing a ShearThickening Fluid with a Helical Ribbon Impeller Type," Chemical Engineering &
Technology, vol. 23, pp. 329-335, 2000.
[134] L. C. Dutto, C. Y. Lepage, and W. G. Habashi, "Effect of the storage format of sparse
linear systems on parallel CFD computations," Computer Methods in Applied Mechanics
and Engineering, vol. 188, pp. 441-453, 2000.
[135] R. Rotondi and C. Arrighetti, "A computational analysis for mass transfer efficiency
optimization in a catalytic converter," International Journal of Modelling and Simulation,
vol. 24, pp. 263-271, 2004.
[136] D. Kandhai, D. J. E. Vidal, A. G. Hoekstra, H. Hoefsloot, P. Iedema, and P. M. A. Sloot,
"Lattice-Boltzmann and finite element simulations of fluid flow in a SMRX static mixer
reactor," International Journal for Numerical Methods in Fluids, vol. 31, pp. 1019-1033,
1999.
[137] W. F. C. Van Wageningen, D. Kandhai, R. F. Mudde, and H. E. A. Van Den Akker,
"Dynamic flow in a kenics static mixer: An assessment of various CFD methods," AIChE
Journal, vol. 50, pp. 1684-1696, 2004.
[138] P. Sagaut, "Toward advanced subgrid models for Lattice-Boltzmann-based Large-eddy
simulation: Theoretical formulations," Computers and Mathematics with Applications,
vol. 59, pp. 2194-2199, 2010.
[139] E. Lorenz, A. Caiazzo, and A. G. Hoekstra, "Corrected momentum exchange method for
lattice Boltzmann simulations of suspension flow," Physical Review E - Statistical,
Nonlinear, and Soft Matter Physics, vol. 79, 2009.
[140] V. Stobiac, P. A. Tanguy, and F. Bertrand, "Boundary conditons for the lattice Boltzmann
method in the case of viscous mixing flows," submitted to Computers and Fluids, 2012.
[141] J. G. M. Eggels, "Direct and large-eddy simulation of turbulent fluid flow using the
lattice-Boltzmann scheme," International Journal of Heat and Fluid Flow, vol. 17, pp.
307-323, 1996.
[142] H. Hartmann, J. J. Derksen, and H. E. A. Van Den Akker, "Macroinstability uncovered in
a Rushton turbine stirred tank by means of LES," AIChE Journal, vol. 50, pp. 2383-2393,
2004.
[143] L. Fradette, G. Thome, P. A. Tanguy, and K. Takenaka, "Power and mixing time study
involving a Maxblend&reg impeller with viscous Newtonian and non-Newtonian fluids,"
Chemical Engineering Research and Design, vol. 85, pp. 1514-1523, 2007.
[144] R. R. Nourgaliev, T. N. Dinh, T. G. Theofanous, and D. Joseph, "The lattice Boltzmann
equation method: Theoretical interpretation, numerics and implications," International
Journal of Multiphase Flow, vol. 29, pp. 117-169, 2003.
[145] M. Junk and Z. Yang, "Convergence of lattice Boltzmann methods for Stokes flows in
periodic and bounded domains," Computers and Mathematics with Applications, vol. 55,
pp. 1481-1491, 2008.
168
[146] A. Caiazzo and M. Junk, "Boundary forces in lattice Boltzmann: Analysis of Momentum
Exchange algorithm," Computers and Mathematics with Applications, vol. 55, pp. 14151423, 2008.
[147] F. Zhao, "Optimal relaxation collisions for lattice Boltzmann methods," In Press,
Corrected Proof, Computers & Mathematics with Applications.
[148] T. Kouda, H. Yano, and F. Yoshinaga, "Effect of agitator configuration on bacterial
cellulose productivity in aerated and agitated culture," Journal of Fermentation and
Bioengineering, vol. 83, pp. 371-376, 1997.
[149] D. Patel, F. Ein-Mozaffari, and M. Mehrvar, "Improving the dynamic performance of
continuous-flow mixing of pseudoplastic fluids possessing yield stress using Maxblend
impeller," Chemical Engineering Research and Design, vol. 90, pp. 514-523, 2012.
[150] O. Hiruta, K. Yamamura, H. Takebe, T. Futamura, K. Iinuma, and H. Tanaka,
"Application of Maxblend fermentor for microbial processes," Journal of Fermentation
and Bioengineering, vol. 83, pp. 79-86, 1997.
[151] N. Dohi, T. Takahashi, K. Minekawa, and Y. Kawase, "Power consumption and solid
suspension performance of large-scale impellers in gas-liquid-solid three-phase stirred
tank reactors," Chemical Engineering Journal, vol. 97, pp. 103-114, 2004.
[152] R. Yatomi, K. Takenaka, K. Takahashi, and P. A. Tanguy, "Mass-transfer characteristics
by surface aeration of large paddle impeller: Application to a polymerization reactor with
liquid level change," Chemical Engineering Research and Design, vol. 86, pp. 13451349, 2008.
[153] A. Tagawa, N. Dohi, and Y. Kawase, "Dispersion of floating solid particles in aerated
stirred tank reactors: Minimum impeller speeds for off-surface and ultimately
homogeneous solid suspension and solids concentration profiles," Industrial and
Engineering Chemistry Research, vol. 45, pp. 818-829, 2006.
[154] A. Fontaine, Y. Guntzburger, F. Bertrand, L. Fradette, and M.-C. Heuzey, "Experimental
investigation of the flow dynamics of rheologically complex fluids in a Maxblend
impeller system using PIV," In press, Corrected Proof, Chemical Engineering Reseach
and Design.
[155] V. Stobiac, P. A. Tanguy, and F. Bertrand, "A thorough investigation of the accuracy of
the extrapolation method for the lattice Boltzmann simulation of viscous fluid flow in a
Maxblend impeller system," submitted to International Journal for Numerical Methods in
Fluids, 2012.
[156] A. Bakker, K. J. Myers, R. W. Ward, and C. K. Lee, "The laminar and turbulent flow
pattern of a pitched blade turbine," Chemical Engineering Research and Design, vol. 74,
pp. 485-491, 1996.
[157] A. Guida, A. W. Nienow, and M. Barigou, "The effects of the azimuthal position of the
measurement plane on the flow parameters determined by PIV within a stirred vessel,"
Chemical Engineering Science, vol. 65, pp. 2454-2463, 2010.
169
[158] K. N. Premnath, M. J. Pattison, and S. Banerjee, "Generalized lattice Boltzmann equation
with forcing term for computation of wall-bounded turbulent flows," Physical Review E Statistical, Nonlinear, and Soft Matter Physics, vol. 79, 2009.
[159] K. N. Premnath and J. Abraham, "Lattice Boltzmann model for axisymmetric multiphase
flows," Physical Review E - Statistical, Nonlinear, and Soft Matter Physics, vol. 71, 2005.
[160] H. Li, C. Pan, and C. T. Miller, "Pore-scale investigation of viscous coupling effects for
two-phase flow in porous media," Physical Review E - Statistical, Nonlinear, and Soft
Matter Physics, vol. 72, 2005.
[161] A. Kuzmin, M. Januszewski, D. Eskin, F. Mostowfi, and J. J. Derksen, "Threedimensional binary-liquid lattice Boltzmann simulation of microchannels with rectangular
cross sections," Chemical Engineering Journal, vol. 178, pp. 306-316, 2011.
[162] J.-H. Yoon and S.-J. Lee, "Direct comparison of 2D PIV and stereoscopic PIV
measurements," Measurement Science and Technology, vol. 13, pp. 1631-1642, 2002.
170
ANNEXE A
– Intégration de LBGK sur l’espace des vitesses
Commençons par définir la moyenne d’un observable
sur l’espace des vitesses
comme :
〈 〉
∫
Si l’on multiplie l’équation (2.5) par
(A.1)
et qu’on l’intègre sur l’espace des vitesses, on
obtient :
〈 〉
〈
〉
〈 〉
〈 〉
(A.2)
avec
〈 〉
Pour
∫
(A.3)
, l’équation (A.2) devient
⃗
De même, pour
(A.4)
, l’équation (A.2) devient
〈
⃗
Si l’on définit le tenseur des contraintes
〉
(A.5)
comme:
∫
⃗
⃗
(A.6)
l’équation (A.5) devient:
⃗
Comme la trace du tenseur
est :
⃗⃗
(A.7)
171
(A.8)
le théorème de l’équipartition de l’énergie permet d’écrire la loi des gaz parfaits :
(A.9)
avec
la pression hydrostatique et
.
172
ANNEXE B
– La décomposition de Chapman-Enskog
Commençons par rappeler l’équation LBGK (2.5) présentée dans la section 2 :
(
(B.1)
)
La décomposition de Chapman-Enskog est basée sur une décomposition des phénomènes en
fonction de leur échelle de temps. Dans cette procédure, la dérivée temporelle est décomposée
comme :
(B.2)
De la même manière, les populations deviennent :
(B.3)
À l’inverse, la dérivée spatiale n’est pas décomposée, c’est-à-dire
Le paramètre
.
est généralement identifié comme étant le nombre de Knudsen
.
Lorsque cette décomposition est appliquée à l’ensemble des termes de l’équation (B.1), il est
possible de les regrouper en fonction de leur ordre en . Pour les ordres 0 et 1, on obtient :
(B.4)
(B.5)